I am not forgotten I am not forgotten I am not forgotten God knows my name
Performance and Cost: Forgotten Variables in Technology Change, Productivity Growth, and Creative...
-
Upload
jeffrey-funk-creating-new-industries -
Category
Business
-
view
631 -
download
7
description
Transcript of Performance and Cost: Forgotten Variables in Technology Change, Productivity Growth, and Creative...
Performance and Cost:
Forgotten Variables in Technology Change,
Productivity Growth, and Creative Destruction
by
Jeffrey Funk
1
National University of Singapore
2
Table of Contents
Chapter 1. Introduction
Chapter 2. The Myth of S-Curves
Chapter 3. The Myth of Slowdowns in Old Technologies Driving Improvements in New
Technologies
Chapter 4. The Myth of Similar Rates of Improvement
Chapter 5. The Myth of Product Design Changes Leading to Increases in Performance and
Process Design Changes Leading to Reductions in Cost
Chapter 6. The Myth of Costs Falling as Cumulative Production Rises
Chapter 7. The Reality of Technology Change
Chapter 8. The Myth of not Being Able to Analyze the Future.
Chapter 9. Final Words
3
Chapter 1, Introduction
Technology change is one of the most powerful forces in the world. It has given us
dramatic improvements in economic productivity and standards of living while at the same
time creating winners and losers at the individual, firm, and country level. Technologies such
as steam engines, electricity, internal combustion engines, automobiles, computers, and
integrated circuits have had a particularly large impact on economic productivity and our
lives. They have given us marvelous new products and services, enabled dramatic changes in
our life styles, and enabled the substitution of machines for humans in an increasing number
of jobs.
The growing importance of technology means that more is expected from technology than
ever before in human history. People expect technology to give us better food, homes, and
vacations, more time for our families, a cleaner environment, a safer workplace, a happier
household, longer lives, more time to be creative, a better sex life, and a fairer society. We
expect scientists and engineers to effectively develop the right technologies, managers to
effectively implement technologies, policy makers to promote good technological solutions,
and universities to help students deal with an increasingly complex world of technology.
These high expectations requires manager, policy makers, and academics to better understand
technology change than is currently done and to understand it at a much deeper level.
Most technology is embodied in complex systems. Complex systems provide us with
food, shelter, water, manufactured products, entertainment, transportation, and work spaces.
These systems can be subdivided into sub-systems each of which can be further subdivided in
what might be called a hierarchy of sub-systems. For example, manufactured products can be
thought of as systems in which components are processed and assembled in global supply
chains. Even the processing of basic components such as steel pipes is a complex system of
4
basic material processes that also rely on the mining and transport of raw materials.
These systems are constantly being reconfigured in response to changes and
improvements in technology. Some reconfigurations occur at lower levels and some occur at
higher levels. Often the reconfigurations at higher levels come from improvements in lower-
level systems. These changes percolate up through hierarchies of sub-systems where changes
at one level impact on the next higher level through decisions made by profit-seeking
managers.
For example, improvements in integrated circuits made possible better computers, mobile
phones, and set-top boxes while improvements in areal recording density made smaller forms
of hard disks and tape players possible. Some firms responded to these improvements with
more innovative electronic products while other firms responded very slowly. The slow
responses by other firms suggest that they did not understand the implications of these
improvements and the changes in higher-level systems that these improvements made
possible.
Sometimes these changes occur over decades and involve not just managers but also
policy makers. For example, the emergence of complex global supply chains occurred as
managers and policy makers adjusted to improvements in information technology, cargo
vessels, ports, rail lines, trucks, containers, and in many other subsystems. These
improvements provided managers and engineers with new choices about how to design
products and services and where to purchase inputs for them. These improvements also
provided policy makers with new choices about regulations, trade rules and their country’s
infrastructure for the global supply chains. The falling cost of information and transportation
enabled products to be global designed and manufactured and the global supply chains
emerged from millions of decisions that were made over many years by managers and policy
makers, each responding to the falling cost of information and transportation.
5
This book is concerned with how these choices emerge from technology change. How
does technology change create new choices for managers, engineers, and policy makers in the
design of systems? How does it lead to changes in the design of systems or to the emergence
of new systems? What technologies are experiencing significant changes and why and thus
what technologies are impacting on higher-level systems? Can we use this information to
more effectively design higher-level systems and to more effectively allocate resources?
Economists often call this change creative destruction while management scholars often
call it innovation. Both terms provide insights, but neither capture the essence of this change.
For creative destruction, although a new system is often created that destroys an old system,
this destruction occurs over many years and through millions of decisions and we are
interested in these micro-changes and decisions that lead to the replacement of old systems
with new ones. Global supply changes emerged from millions of decisions that were made in
response to the falling cost of information and transportation technology over many years.
Innovation is also too general and abstract for what we are trying to describe. The term
does not help us understand the millions of micro-changes and decisions that led to the
emergence of the global supply chains. While some of these decisions were more innovative
than others (e.g., container shipping), a focus on such innovations does not help us
understand the impact of improvements in lower level systems on higher level systems. For
example, the emergence of global supply chains primarily came from incremental
improvements in information systems, improvements in information systems primarily came
from incremental improvements in computers, and incremental improvements in computers
primarily came from incremental improvements in integrated circuits. Which improvements
were innovations and which were not? Identifying the impact of improvements in lower level
systems on higher level ones helps us better understand how higher level systems change than
does identifying which of the changes qualifies as an important innovation, a not so important
6
innovation, or not an innovation at all.
Management scholars will usually focus on who captures the most value from innovations
and thus the types of strategies that lead to high profits. This book is less interested in these
strategies and more interested in the front end of this process: how do choices about new
systems become available and thus how can managers, engineers and policy makers find new
choices and effectively consider then? How does technology change cause new systems to
become economically possible and thus provide managers with new choices about designs
and strategies?
One premise of this book is that any discussion of new systems and how they emerge
from technology change must focus on performance and cost. They are important because
managers and policy makers are concerned with their absolute levels and because they
respond to changes in them. Managers are concerned with absolute levels of performance and
cost and changes in them because they impact on their value propositions to customers and
their profits. One of the ways they respond to changes in the performance and cost of inputs
is by changing the ways systems are designed and configured.
Policy makers also respond to both changes in performance and cost and to their absolute
levels. Some policy makers are more concerned with the former while others are more
concerned with absolute levels. Absolute levels of performance and cost are important
because they directly impact on productivity, standards of living, and quality of life, issues
that often concern voters in the long term. For example, without increases in crop yield (e.g.,
bushels of wheat per acre of land) and labor productivity (output per hour of farm workers) it
would be difficult to feed the world’s seven billion people and to keep doing so as the world
population increases. Furthermore, the challenges of improving these performance and cost
indices are exacerbated by water shortages and demands for lower use of pesticides,
insecticides, and fungicides. In other words, meeting these new challenges require us to
7
increase crop yields and productivity with respect to an increasing list of inputs.
Almost every problem can be represented by these types of output-to input ratios, of which
examples are shown in Table 1.1. Better mobile phones have higher faster speeds (bit per
second), higher spectral efficiencies (bits per second and bandwidth), more storage capacity
(bits per chip), better cameras (number of pixels) and faster microprocessors (cycles per
second), which may occur through parallel processing (number of cores). Better computers
are measured in terms of instructions per unit time, per unit-cost, or per kw-hour. More cost
effective DNA sequencing and synthesizing require lower sequencing or synthesizing per unit
cost.
Addressing climate change requires improvement in energy output per weight of carbon
emissions along with more efficient and cost effective solar cells. The cost of solar cells is
measured in terms of cost per peak Watt. Energy efficient lighting and displays require higher
luminosity per Watt light bulbs such as LEDs (light-emitting diodes). Economically effective
electric vehicles require higher energy and power storage densities from batteries or
alternatives such as capacitors and flywheels.
A second reason for our interest in performance-to cost ratios is that new technologies and
systems composed from new technologies must reach specific levels of cost and performance
for them to become economically feasible. For example, the cost of solar cells on a per-kw-
hour or a per-peak Watt level must fall to a certain levels before they will begin to diffuse.
More generally speaking, technologies with low carbon emissions must also have much
lower costs than they currently do in order for them to become economically feasible without
large subsidies and contribute towards reductions in carbon emissions. Thus, a major
challenge for clean energy advocates is to find those technologies that are likely to experience
cost reductions than are other technologies and become economically feasible in the near
future.
8
But this raises a number of questions. What drives improvements in a technology’s cost
and performance? Is it improvements on the factory floor, better components, or new product
and process designs? And why do some technologies experience rapid rates of improvement
than do other technologies? Is it because of greater production, demand, R&D, or something
else? Unless we understand the answers to these questions, it is difficult to understand
creative destruction and innovation, to address climate change, and to help engineers and
other young entrepreneurs create
9
Table 1.1 Examples of Output-to Input Ratios that are Important to Humanity
Technology
Domain
Measure of Performance
or Sub-Technology
Examples of Output-to Input Ratios
Agriculture Productivity of land,
people, water, and
fertilizers
Crop production per acre of land, per person-
hour, per water volume, and per volume or
mass of chemicals and fertilizers
Housing Floor space Area per person
Electricity
Generation
Cost, efficiency Cost per kw hour of electricity, energy per
weight of carbon emissions, energy generated
per area of land, cost per peak Watt of Solar or
Wind, efficiency of solar cells, cost of carbon
sequestration of fossil fuels
Electricity
Transmission
Cost, efficiency Cost per distance of transmitting electricity,
energy loss per length of transmission line,
current x length or current x length per dollar
for transmission materials such as
superconductors
Electricity
Storage
Cost, efficiency Energy stored (e.g., joules) per mass (e.g., kg),
volume (e.g., liters), or cost; power (watts)
generated per mass, volume, or cost
Transportation Cost, efficiency Cost per weight-mile for transporting cargo,
cost per person-mile of transporting humans,
amount of energy needed to transport cargo
(weight) or humans per mile
Lighting and
Displays
Lighting and Displays Luminosity per Watt, lumens per dollar
Displays Square meters per dollar
Lasers Power density, cost per Watt
Health Care Quality of life Longevity and quality of vision, hearing, smell,
other organs, and of mobility
Cost Percentage of GNP devoted to health care, cost
10
of specific operations and procedures
Information
Processing
Microprocessor ICs Number of transistors per chip, cost per
transistor and cycle time, number of cycles per
second
Camera chips Pixels per dollar, light sensitivity
new systems. Evidence of a lack of understanding can be found in the emphasis on wind
turbines and batteries, both of which are experiencing very slow rates of improvement. To do
this, however, we must return to the basics and address the assumptions that form the basis
for most management and economic analyses of technology. In reassessing these
assumptions, we have found that many of them are very wrong and thus we call them
“myths.” These myths distort the reality of technology change and mislead the choices
available for engineers, scientists, managers, and policy makers. Our empirical research has
identified six myths that impact strongly on how these people think about technology change:
#1 Performance vs. time curves resemble an S-curve
#2 Slowing rates of improvement in old technologies drive the development of new
technologies
#3 All technologies have the potential for rapid rates of improvement.
#4 Product design changes drives performance increases and process design changes
drives cost reductions, with product preceding process design changes in a
technology’s life cycle.
#5 Costs fall as cumulative production rises in a learning curve
#6 The future of new technologies cannot be analyzed.
These myths are largely based on metaphors and anecdotal evidence that were presented
11
decades ago and that have not been systematically re-examined. The one exception is the
learning curve, but even here the empirical evidence has been selective, ignoring the
improvements that occur for most technologies before commercial production begins. This
book summarizes the empirical research that proves these myths wrong and it replaces them
with much more accurate descriptions of reality. These more accurate descriptions of reality
suggest more appropriate methods of creating new systems that are very different from the
ones based on the myths.
The first two myths prevent us from understanding the shapes of performance or cost vs.
time curves and thus rates of improvement for what Giovanni Dosi calls “technology
trajectories.” The myth of S-curves (See left side of Figure 1.1), which is explored in Chapter
2, makes us believe that rates of improvement will strongly accelerate when demand, R&D
spending or something else increases and also that slowdowns and physical limits will
quickly emerge. The first half of this myth facilitates the exaggeration and hype of new
technologies and makes it appear as if any technology is right around the corner as long as
governments, firms, and other potential investors make the right investment and aren’t too
focused on the short term. As one of my colleagues said as I discussed the slow rate of
improvement for electric batteries, “but improvements will accelerate once the demand
increases.” In other words, real data isn’t as important as a belief in the accelerations that
form the basis for the S-curve. Purported accelerations have been used to justify investments
in solar cells, wind turbines, and electric vehicles1.
12
The second half of this myth makes us believe that slowdowns and limits will quickly emerge
and thus we must quickly jump to a new technology. Michael Tushman and Philip Anderson2
make the strongest argument for steep accelerations that are quickly followed by slowdowns
and limits in that long periods of incremental change are punctuated by relatively rare
innovations that provide sharp price-performance improvements over existing technologies.
Or in their words, “technology evolves through relatively long periods of incremental change
punctuated by relatively rare innovations that radically improve the state of the art.” Building
from this paper and findings from evolutionary biology, other scholars use the term
“punctuated equilibrium” to describe a long-term process in which stable environments are
punctuated by new technologies that provide jumps in performance and that also instigate
organizational transformation. Punctuated equilibrium is one of the most powerful and often
cited theories in management and it increases the hype of S-curves3.
Unfortunately, there is little evidence for punctuated equilibrium or S-curves particularly if
one looks at logarithmic plots of data from engineering and science journals. One must use
13
logarithmic plots because a straight line on a logarithmic plot means constant rates of
improvement over time while a straight line on a linear plot means a decreasing rate of
improvement. Engineering and science journals plot data for rapidly improving technologies
on logarithmic plots and such plots typically include some combination of best laboratory
results and commercialized products. Without data from best laboratory results, there may be
gaps in the data that come from time periods between product releases and that falsely imply
jumps in performance4.
Our analysis of performance vs. time curves for 25 technologies and 32 measures show
that the shape of performance-versus time curves more closely resembles a straight line on a
logarithmic plot (See right side of Figure 1.2) than the classical S-curve and that punctuated
equilibrium is a highly misleading metaphor. Thus, we should not expect or plan for early
accelerations or for early limits. Instead, we should plan for and expect incremental
improvements to occur at fairly constant rates of improvement. These constant rates of
improvement allow us to compare rates of improvements for different technologies and to
consider these comparisons when long term plans for development and implementation are
being considered. Firms can and do consider rates of improvement when they choose
research projects and clean energy advocates should consider rates of improvement when
they propose potential solutions.
The second myth, which is explored in Chapter 3, extends the myth of S-curves further
into a fantasy land in which slowdowns in a single old technology are perfectly linked with
accelerations in a single new technology (See left side of Figure 1.2). In other words, it is the
slowdown in an old technology that causes the development and improvement of a new
technology and for an acceleration in the rate of improvement of this “single” new
technology to occur. This myth about “linked S-curves” suggests that demand is the most
important lever for policy and it is the most important driver for firms when they create R&D
14
strategies. According to this myth, a slowdown in an old technology represents demand for a
better solution and the acceleration in the new technology represents the response to the
slowdown.
Since accelerations were not found in our analysis of the first myth of S-curves, there is
little chance that such accelerations will occur at the same time as slowdowns in the linked S-
curves shown in Figure 1.2. The best one can hope for with this myth is that a slowdown
occurs in an old technology as a new technology experiences some form of improvements.
Chapter 3 analyzes this possibility using 15 pairs of old and new technologies and finds only
two technologies experienced statistically significant slower rates of improvement after a
performance metric was recorded for a new technology, thus suggesting that slowdowns do
not generally cause the development and improvement of new technologies. Furthermore,
Chapter 3 finds that multiple new technologies are being developed simultaneously as
replacements for an old technology (See right side of Figure 1.2), unlike the single
performance-to cost curve that is presented in the linked S-curve theory (left side of Figure
1.2).
15
Chapter 3’s analysis is also more consistent with the “technology push” than of “market
pull” theories of technology change, a debate that has continued for more than 40 years with
Clayton Christensen’s theory of disruptive innovation providing recent evidence for market
pull. Proponents of technology push argue that technology change is driven by universities
and other laboratories that “push” technologies into the marketplace while others argue that
technology change is driven by a market that “pulls” ideas for new technologies from the
minds of scientists and engineers in universities and laboratories through targeted research.
One of the persuasive arguments for technology push in the 1970s was that few of the needs
addressed by the important “innovations” of the 20th century had been recognized before the
new technology was developed. Looking at subsequent innovations such as cellular phones
and the Internet, one can also perceive this lack of early recognition in the needs for these
innovations.
Chapter 3’s analysis suggests that performance vs. time curves are being driven much more
by supply than demand-side factors. Although demand becomes more important as a new
16
technology experiences improvements and thus nears economic feasibility, the initial
improvements have more to do with supply side factors such as advances in science or
accidental discoveries than with a slowdown in an old technology. It is only as new
technologies are developed and improved that the state of the old technology becomes a more
relevant driver of the improvements in the new technologies.
The third myth, which is explored in Chapter 4, is that all technologies have the same
potential for improvements and thus the same potential for rapid rates of improvement.
Although this myth has not been explicitly stated anywhere to our knowledge, it is implicitly
stated almost everywhere. One type of evidence for the existence of this myth is that rates of
improvement are rarely discussed and few people are aware of them, one exception is
Moore’s Law. An ignorance of rates of improvement is particularly common among social
scientists even those who specialize in technology. For example, Clayton Christensen’s books
and papers on disruptive innovations do not mention rates of improvement even though rates
of improvement clearly impact on the replacement of old technologies with new ones and
even though his theory is largely based on an analysis of hard disk drives, which have
experienced unusually rapid rates of improvement. Christensen conveniently ignores such
details perhaps in an attempt to promote the universality of his theory.
Ignoring rates of improvement is also common among advocates of sustainability. Most
papers, books, presentations, and courses on sustainability rarely mention rates of
improvement and the particularly slow rates of improvement for wind turbines and batteries.
A notable exception is Vaclav Smil’s books on energy and sustainability but his books sell far
fewer copies than do those of Thomas Friedman, who does not mention rates of improvement
in his books or in his New York Times columns on clean energy. Even discussions with
professors of sustainability or sustainability-related organizations such as the
Intergovernmental Panel on Climate Change or the Society of Risk Analysis will merely
17
produce blank stares, laughs, and sometimes anger. A common response is “it’s not just about
technology,” as if considering rates of improvement prevents ones from considering other
factors.
Equally troubling, the few analyses of rates of improvement confuse rates of improvement
for each doubling of cumulative production with annual rates of improvement. This causes
them to overestimate the rates of improvement for some technologies since annual rates and
those for doubling will only be the same when cumulative production is doubling each year,
which rarely occurs. Most technologies experience much slower rates of production growth
than 100% per year and thus their annual rates of improvement are 1/10 to ¼ the rates for
each doubling of cumulative production.
-10 -8 -6 -4 -2 0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40>420
5
10
15
20
25
FIgure 1.3. Number of Technologies by Annual Rates of Improvement
Annual Rates of Improvement
The reality is that some technologies have much faster annual rates of improvement than
do other technologies. As shown in Figure 1.3, annual rates of improvement range from
minus 10% to some greater than 42%. Two-thirds of the technologies have rates of less than
9% and 89% of them have rates of less than 15% per year. Furthermore, the empty space
between 15% and 25% and our analysis of chemical vs. non-chemical technologies suggest
18
that there are multiple distributions in Figure 1.3, which enable us to draw conclusions about
how these improvements occur. In combination with an analysis of production volumes vs.
rates of improvement, we are able to conclude that increases in production scale have a much
smaller impact on improvements in cost and performance than do changes in product and
process designs. After analyzing other possible reasons (greater demand and R&D), for the
differences in annual rates of improvements, Chapter 4 concludes that some technologies
have greater opportunities for improvements from product and process designs than do other
technologies. This conclusion was reached by Nathan Rosenberg more than 40 years ago but
contradicts the recent emphases on demand-based theories (e.g., Clayton Christensen’s theory
of disruptive innovation) and on demand-based subsidies as a tool for promoting clean energy
technologies.
The fact is that technologies with rapid rates of improvement generally have had and will
continue to have a larger impact on our world than do other technologies. Technologies such
as integrated circuits (ICs), hard disk drives, lasers, and computers have had a large impact on
our world partly because they have experienced one or two orders of magnitude
improvements each decade resulting in 5 to 10 orders of magnitude improvements over the
last 50 years. They have enabled new forms of systems to emerge, one of which is the
Internet and another of which is the global supply chains that were mentioned earlier and that
depend on the Internet. They have also enabled other systems to emerge, many of which are
defined as disruptive innovations by Clayton Christensen. Thus, if we are trying to
understand technology change or we are looking for disruptive innovations, we should be
focusing on technologies with rapid rates of improvement as this will provide us with better
ideas than will looking for Christensen’s low-end innovations.
Technologies with rapid rates of improvement in performance and/or cost will also have
faster rates of diffusion than will ones with slower rates of improvement. While rates of
19
diffusion depends on a variety of factors, the largest factor identified in diffusion studies for
the U.S. is the profitability of a technology for users5. The users that profit most from a new
technology, whether these users are industrial ones or consumers, are the first adopters and
faster rates of improvement in performance and/or cost mean faster rates of improvement in
the profitability for users. One reason for the large importance of profitability for users in
determining rates of diffusion is that the U.S. and other developed countries have less barriers
to entry for “creative destruction” than do other countries, which is of course why the U.S.
and other developed countries are richer than are poorer ones.
In other words, slow rates of improvement are a type of implementation problem that must
be overcome if a technology is to succeed. Scholars of technology management often focus
on implementation problems and thus argue that rates of improvement are less important than
are organizational or regulatory challenges. This book argues that slow rates of improvement
are probably a larger problem than are organizational and regulatory challenges in developed
countries because these countries have become good at solving organization and regulatory
problems through existing and new firms, venture capital, and changes in regulations by
policy makers.
The importance of slow rates of improvement as an implementation problem means that
understanding why some technologies have more rapid rates of improvement than do others
is a critical question. Understanding the answer to this question can help us achieve more
rapid rates of improvement in existing technologies and more importantly identify
technologies with a greater chance of achieving rapid rates than do other technologies.
Answering this question about the reasons behind rapid rates requires us to address two other
myths and both these two myths and the prior two myths also contribute to the myth that all
technologies have the same potential for improvements. When one believes in linked S-
curves, accelerations occurring early in a technology’s lifecycle, the slowdowns that drive
20
these accelerations and two other myths, it is easy to believe that all technologies have the
same potential for improvements.
The fourth myth, which is explored in Chapter 5, is that product design changes lead to
increases in performance early in the life cycle and process design changes lead to reductions
in costs later in the life cycle (See Figure 1.4). This myth causes firm strategies and
government policies to focus on product design changes to achieve performance increases
and to focus on process design changes to achieve cost reductions. “Process” in general may
include any activities that contribute to cost and the strategic importance of process
innovation for competition increases through the life cycle of an industry, as the opportunities
decline for product innovations that lead to improvements in functional performance. Many
scholars cite some version of this assessment and some take this assessment further and argue
that increasing returns to process R&D lead to a shakeout, i.e., dramatic decline, in the
number of firms6.
This myth is addressed by analyzing the cost and performance of 17 different technologies,
21
22 unique time-series pairs of performance and cost, 358 unique data points and 705 total
years of data for these time-series pairs. Technologies such as chemicals, agricultural
products and materials are excluded since they often have a fixed chemical composition and
thus improvements in performance do not occur and most of their improvements in cost are
driven by process innovations or by increasing the scale of their production equipment7. But
for other technologies, our analysis shows that improvements in performance and cost for
most products are highly correlated over many decades suggesting that the same design
changes are impacting on improvements in both cost and performance.
What types of design changes might cause simultaneous improvements in both cost and
performance and also cause very different rates of improvement in different technologies?
There are several possibilities and this and the following chapters explore them. One
possibility is that product and process design changes do drive improvements in performance
and cost respectively and that firms are implementing product and process innovations at the
same time because they want to simultaneously reduce costs and improve performance. A
second possibility is that some changes in product design lead to both improvements in
performance and cost. For example, improvements in the efficiency of solar cells, lights and
displays and in the speed of electronic devices, computers and other electronic products also
lead to reductions in the cost of these technologies. A third possibility is that inter-related
product and process design changes are implemented by either single firms in what may be
called “integral” design or multiple firms in what many call modular design. Our analysis of
the next myth, the fifth one, helps us better understand the specific types of inter-related
product and process design changes that lead to improvements in cost and performance and
explains why some technologies experience more rapid rates of improvements than do other
technologies.
The fifth myth, which is analyzed in Chapter 6, is that costs fall as cumulative production
22
rises and as improvements are made to processes on the factory floor (See Figure 1.5). Since
the publication of Theodore Paul Wright’s analysis of fighter jet costs in 1936, empirical
analyses correlating cost reductions to cumulative production have grown extensively in what
some call learning curves. The early work on learning curves was mostly done on single
designs in specific factories and thus analyzed the impact of factory level changes on factory
productivity. Subsequently, learning curves have been applied to technologies that are
manufactured with new designs and in new factories where the output variable might be cost
or performance, albeit these models are now often called experience curves. For example, the
costs of ships, solar cells, semiconductor memory, chemicals, primary metals, and food have
been analyzed using this approach, across significant design changes and often throughout all
global factories8.
Linking cumulative production to reductions in cost or improvements in performance can
lead to confusion about how the improvements in cost and performance are being achieved.
Are these improvements being achieved through changes made on the factory floor or
23
through changes made in laboratories to product and process designs? Some believe that an
empirical linkage being cost and cumulative production suggests most of the improvements
are occurring on the factory floor while others note that cumulative production indirectly
leads to improvements in performance or cost. Increases in production are linked with
expected future production and lead to increased incentive to perform process-related and
product R&D9 where the results of the increased R&D spending lead to improvements in
performance or cost. This argument is also implicit in Christensen’s10 analysis of hard disk
drives, computers and other “disruptive” technologies in that the emergence of a low-end
product lead to increases in R&D spending and thus rapid improvements in the new product,
which in turn leads to replacement of the dominant technology by the low-end innovation.
Linking cumulative production to reductions in cost or improvements in performance
also makes it easier for policy makers to justify demand-based subsidies for desirable new
technologies such as clean energy ones. It is particularly easy to justify demand-based
subsidies if the cost reductions are coming from activities on the factory floor and thus the
subsidies will probably encourage the cost reductions. However, what if they are not coming
from activities on the factory floor but instead are coming from R&D activities in the
laboratories? If it is the latter, then a different set of policies are needed to achieve cost
reductions than the current emphasis on demand-based subsidies.
Chapter 5 and 6’s analysis suggests a different reality than the myth of increases in
cumulative production lead to cost reductions. The correlation that is found between
improvements in cost and performance in Chapter 5 suggests that costs are falling because of
changes in product and process design that also impact on improvements in performance.
Thus, increases in cumulative production and changes in process design by themselves
cannot be the reasons for falling costs. This argument is taken one step further in Chapter 6
by conducting an analysis of technologies that experienced rapid rates of improvement with
24
no commercial production. These include organic materials for transistors, solar cells, and
light-emitting diodes (LEDs), quantum dots for displays and solar cells, new forms of non-
volatile memory integrated circuits such as resistive or magnetic RAM (random access
memory), carbon nanotubes, superconducting materials for both energy transmission and
integrated circuits, and quantum computers.
Since there was no commercial production during most of the time periods in which the
rapid improvements occurred, the improvements must be from factors other than factory floor
activities and this chapter examines the specific product and process design changes that
enabled the rapid improvements to occur. Our analysis suggests that these design changes
occur in laboratories and there are primarily two types of inter-related product and process
design changes: 1) creating new materials (and their associated processes); and 2) reducing
the scale of features in the product through process improvements.
Chapter 7 presents a reality of technology change that is based on the analyses from the
previous chapters. Performance vs. time curves more closely resemble straight lines on a
logarithmic plot than an S-curve. Accelerations do not occur and physical limits take decades
if not longer to emerge. The improvements in the new technology are not driven by a
slowdown in the old technology and in fact multiple technologies are competing to replace a
new technology even without a slowdown in the old technology. This competition primarily
occurs in laboratories where advances in science and accidental discoveries determine the
timing of the first recorded performance metrics and the rates of improvement largely
determine the eventual winning technologies. A lack of S-curves means that pre-production
early rates of improvement can provide an important signal for future rates and thus help us
identify the technologies with the large potential for rapid improvements.
Different technologies have different rates of improvement primarily because of
differences in opportunities for improvements. Simply put, some technologies experience
25
more rapid rates of improvement because they benefit more from the two inter-related
product and process design changes mentioned above than do other technologies. These
rapidly improving technologies have a greater chance of becoming economically feasible
than do other technologies for either the first application or an increasing number of
applications and they have a higher chance of impacting on higher-level systems.
The impact on higher-level systems may be the most important impact since most
technologies are embodied in complex systems. As noted earlier, complex systems provide us
with food, shelter, water, manufactured products, entertainment, transportation, and work
spaces. These systems can be subdivided into sub-systems each of which can be further
subdivided in what might be called a hierarchy of sub-systems. If a system contains a
technology that experiences rapid rates of improvement and that technology strongly impacts
on the system’s cost and performance, the system will also likely experience rapid rates of
improvement.
This reality of technology change is described in two ways. The first way is to show how
the two types of product and process design changes impact on a large number of
technologies, including their impact on higher-level systems. The second way is to discuss
technology change from the standpoint of invention and commercialization. How do new
technologies proceed from invention to commercialization and what does this mean for
universities, startups, and other firms?
Chapter 8 uses this better reality of technology change to addresses the sixth myth: we
can’t analyze the future. This is probably the most controversial and emotional myth because
most social scientists dismiss analyses of the future as useless while failing to recognize that
managers, policy makers, professors, and students are constantly making decisions that imply
predictions about the future. When managers introduce R&D budgets or new product or
students choose university majors, they are making a prediction about the future. The firms
26
are predicting that the R&D budgets and new products will probably provide greater benefits
than their costs. The students (or their parents) are predicting that the university degrees will
provide them with benefits that are greater than the costs of the education. Even social
science professors, who enjoy criticizing predictions, are implicitly making a prediction when
they draw conclusions about specific policies, strategies, or organizational designs. By
concluding that something has worked better than something else, they are implying that this
conclusion will be true in the future.
More importantly, criticizing predictions prevents social scientists from understanding the
challenges of searching for entrepreneurial opportunities, solving global problems such as
sustainability, helping students search for opportunities and address sustainability, and setting
effective R&D policies. Where should young entrepreneurs search for opportunities? How
should they try to address sustainability? Which technologies and markets should they
consider? What kinds of systems should they introduce? These types of questions must be
addressed in order to search for entrepreneurial opportunities, solve global problems such as
sustainability, help students search for opportunities and address sustainability, and set
effective R&D policies. Unless social scientists help their students make such decisions, they
cannot help them address these issues and helping them make such decisions requires them to
make predictions about which solutions have the largest chance of success.
An aversion to helping young entrepreneurs choose the technologies or markets to look for
opportunities reminds me of a joke that economists like to tell. A man walking down the
street sees a man crawling around on his hands and knees under a lamp. The pedestrian asks
the crawler, what are you looking for. The crawler says “my keys. I lost them over there, as
he points to a place not under the lamp.” The pedestrian then asks while pointing at the place
where he is crawling, why are you looking there?” The crawler says, “this is where the light
is.”
27
Analyzing the future requires us to look in places where the light is not very bright.
Although the previous chapters of this book have increased the luminosity of the light by
disproving myths and replacing them with more useful insights into of technology change,
the future is still dimly lit. It is all about probabilities. Which types of systems have the
largest probability of becoming economically feasible? Searching for entrepreneurial
opportunities, solving global problems such as sustainability, helping students search for
opportunities and address sustainability, and setting effective R&D policies requires us to
think about the technologies with the greatest chance of achieving improvement and thus the
decisions that will probably lead to the best futures.
There are two key factors in an analysis of when new technologies might become
economically feasible and they are summarized in Figure 1.6: the extent to which
improvements are needed and the annual rate of improvement. Both impact on when new
technologies might become economically feasible, when they might do so for a larger number
of applications, or when they might impact on higher-level system. This book is primarily
concerned with the ones in the right side of Figure 1.6 since these are the technologies that
are often ignored and misunderstood.
28
Chapter 8 distinguishes between predictions of the future and analyzing the future. After
analyzing predictions made by MIT’s Technology Review, it then shows that this book’s
insight about rapid improvements provide us with a better way to analyze the future than does
Technology Review’s method of expert elicitation. Not only do these rapid rates of
improvement provide us with better predictions than does Technology Review’s method, it
also enables us to reduce forecast error by gathering and analyzing more data. The other
chapters reduce forecast error by helping us better understand technology change. This
chapter focuses on reducing forecast error by better understanding rates of improvement,
comparisons between old and new technologies, the composition of systems, and the extent
to which improvements are needed.
This chapter draws on research done by the author for a course on technology change. This
course helps students understand the composition of existing and potential systems,
comparisons between old and new systems, rates of improvement that impact on these
systems, and the extent to which improvements are needed before these new systems become
29
economically feasible. Effectively analyzing these factors and the uncertainty in them can
help students better understand when new technologies become economically feasible. More
specifically, better understanding the composition of systems can lead to better cost and
performance estimates, better comparisons between new and old systems, and better
estimates for when the new systems might become economically feasible. This is one goal of
this book and has been the goal of the author for more than 10 years.
Chapter 2
The Myth of S-Curves
30
S-curves for performance vs. time are one of the most widely accepted theories of
technology change. They are widely cited in academic papers11 and in general books on
technology change12. Some13 argue for highly accelerated S-curves in which long periods of
incremental change are punctuated by breakthroughs that provide sharp price-performance
improvements over existing technologies14.
However, these papers or books provide little empirical evidence for punctuated
equilibrium or S-curves on a logarithmic plot or even little explanation for why performance
vs. time curves might resemble an S-curve particularly during the early part of the purported
S-curve when an acceleration is supposed to occur15. The purported theory is that
improvements accelerate as firms and government agencies move research funds from an old
to a new technology in response to increases in demand for the new technology or to a
slowdown in the rate of improvement in the old technology16. The acceleration may also
occur as the technology is better understood by scientists and firms, constraints are overcome,
and complementary technologies developed and implemented17. For the later part of the
purported S-curve, the rates of improvement slow as the cost of marginal improvements
increases and natural limits emerge; this causes research funds to move to a still newer
technology and thus the newer technology’s rate of improvement begins to accelerate18.
This chapter shows that performance vs. time curves more closely resemble straight lines
on a logarithmic plot, than the classical S-curves and it describes the rational for these
straight lines. We emphasize logarithmic plots because straight lines on logarithmic plots
represent constant rates of improvement while straight lines on linear plots represent
decreasing rates of improvement19. If Moore’s Law or similar data is plotted on a linear scale,
the plot would show a very slow increase in the numbers of transistors per chip until these
numbers approached the current order of magnitude figure and then a sharp acceleration
would “appear” in the numbers of transistors per chip. We put the word “appear” in quotes
31
since as some have noted20, there is no acceleration because the percentage increases each
year are constant in a logarithmic plot; it is just that linear plots do not accurately show
exponential change. Another misleading result of plotting rapid rates of improvement on a
linear plot is that the location of the acceleration is always about five-to ten years before the
most recent data point thus causing the illusory acceleration to depend on the range of the
data analysis. Clearly, this would be a nonsensical result since the location of the acceleration
should be independent of the last data point plotted.
This chapter uses time-series data from engineering and science journals to analyze the
shapes of performance vs. time curves. The technologies are summarized in Table 2.1 and the
methodology is described in more detail in the Appendix. Engineers and scientists typically
plot performance data on a logarithmic plot using a combination of best laboratory results and
commercialized products, with data from best laboratory results often used during the early
years of the technology before commercial products are regularly introduced. Without data on
best laboratory results, there may be gaps in the data that come from time lags between
product releases and that falsely imply jumps in performance21. We return to this issue in the
discussion section of the chapter.
2. The Evidence
Figure 2.1 (in Appendix) is a meta-figure that includes 25 smaller figures of which several
include multiple technologies in order to conserve space. None of the 32 time-series curves
display the classical S-curve. The second half of an S-curve, i.e., limits, is only evident in one
technology, the best laboratory efficiency of amorphous silicon solar cells (Figure 2.1.c), and
this is consistent with the statistical analysis in this chapter’s appendix (negative coefficient
of the time squared variable in the log model and the low p-value of less than 0.0001). A
reduced rate of improvement in efficiency is to be expected since there is a maximum
Table 2.1 Technologies with Recent Rapid Rates of Improvement
32
Technology
Domain
Sub-Technology Dimensions of measure Time Period Improvement
Rate Per
Year
Energy
Trans-
formation
Light Emitting
Diodes (LEDs)
Lumens per package, red 1968-2005 38%
Lumens per package,
white
2000-2010 95%
Lumens per dollar, red 1973-2005 28%
Solar
Cells
Crystal
Silicon
kwH per dollar 1957-2003 16%
Organic Efficiency 2001-2012 11%
Amorpho
us Silicon
Efficiency 1976-2013 8.7%
Energy
Trans-
mission
Superconducting
Cables
Current (amps) x length
(meters) - YBaCuO
2002-2011 53%
Energy
Storage
Li-ion Batteries Energy (joules) per
volume (cc)
1986-2000 8.1%
Information
Trans-
formation
Logic Chips Number of transistors
per chip
1959-1975 91%
Microprocessors 1971-2011 40%
Dynamic RAM 1970-2004 44%
33
Bits per dollar 1971-2010 43%
Flash Memory 1992-2007 66%
Camera chips Pixels per dollar 1983-2013 49%
Light sensitivity 1986-2008 18%
Pixel size 1986-2008 21%
Organic
Transistors
Mobility cm2 per Volt-
second
1984-2007 94%
Computers Instructions per unit time
and dollar
1945-2008 38%
Computations per kw-hr 1947-2009
Computer
Tomography
1/(scan time x
resolution)
1971-2006 62%
Information
Storage
Mag-
netic
Storage
Disks Areal density 1956-2007 37%
Tape Volumetric density 1952-2000 24%
Bits per dollar 1952-2004 30%
Ferro-
electric RAM
Storage Capacity 2001-2009 37%
Magneto RAM Storage Capacity 2002-2011 58%
Phase Change
RAM
Storage Capacity 2004-2012 63%
34
Information
Trans-
mission
Glass Fiber Km per decibel 1966-1983 72%
Last Mile
Wireline
Bits per second 1982-2010 49%
Wireless
(Cellular)
Bits per second 1895-2008 18.7%
1996-2013 79%
Biological
Trans-
DNA Sequencing Base Pairs per dollar 1971-2003 78%
Cellulosic
Ethanol
Output (liters) per dollar 2001-2012 14%
Biological
Trans-
portation
Aircraft Number of Passengers
Times Speed
1926-1975 13%
RAM: random access memory.
Sources: (Haitz and Tsao, 2110; Nemet, 2006; NREL, 2013; Shiohara et al, 2013;
Selvamanickam V 2011; Koh and Magee, 2008; Computer History, 2014; Singularity.com
2014; iNEMI, 2010; ICKnowledge, 2009; futurefab, 2013; Suzuki, 2010; Hasegawa and
Takeya, 2009; Koomey et al, 2011; Koh and Magee, 2006; Kalender, 2006; Yoon, 2010; Koh
and Magee, 2006; ISSCC, 2013: NAS/NRC, 1989; Brown, 2011; ISSCC, 2013; NHGRI,
2014; Singularity.com 2014; Service, 2013; Martino, 1971)
35
theoretical efficiency for amorphous solar cells and the current values may be approaching a
limit that is consistent with the maximum theoretical efficiency. However, if we had data on
output (kwHours) per dollar for amorphous silicon cells, we would expect this dimension of
performance to continue rising for many years, similar to what has happened with output per
dollar for crystalline silicon solar cells (See Figure 2.1.c.). As an aside, we note that another
data base22 shows that crystalline silicon solar cells have continued to experience rapid
reductions in cost beyond 200323, which is the last data point in Tables 2.1 and 2.2.
The first half of an S-curve, i.e., acceleration, is also only evident in one technology,
cellular telecommunications (Figure 2.1.u). The measure of bits per second is relatively
constant early in the time series followed by a steep acceleration in the late 1990s. This is
also consistent with the statistical analysis in this chapter’s appendix (positive coefficient of
the time squared variable in the second-order log model and the low p-value of 0.03). This
acceleration is also to be expected since cellular phones were first used for voice
communication and data speeds only became important in the late 1990s as the performance
of complementary technologies such as displays reached the levels necessary for data speeds
to become important; this explanation is consistent with the theory of S-curves24.
Nevertheless, a more appropriate measure of performance for the early years of cellular
phones would probably be number of voice conversations possible per unit of spectrum and
this measure of performance might follow a straight line on a logarithmic plot. Although a
data set on conversations per unit of spectrum vs. time displays a straight line on a
logarithmic plot25, the density of data points is too small to include in Table 2.2.
In addition to wireless cellular and amorphous silicon solar cells, several other curves
deviate slightly from a straight line on a logarithmic plot because these curves include
slowdowns and accelerations either in the middle or near the ends of the time series. The
statistical analysis for two of these technologies are consistent with the statistical analysis in
36
this chapter’s appendix in that they have coefficients for a time squared variable that are both
negative and significant. They are organic transistors (Figure 2.1.k) and computer-
tomography (Figure 2.1.n). Although there is not enough data points for the coefficient to be
significant, Figure 2.1.r suggests that the rate of improvement may also have slowed for
MRAM and FeRAM. Some might interpret these slowdowns as the inevitable move towards
limits and thus evidence of the latter part of an S-curve.
We urge caution in making such an interpretation for several reasons of which one will be
emphasized here. Several other technologies in Figure 2.1 experienced accelerations in the
rates of improvement in the middle or near the end of their time series and these accelerations
followed the emergence of slower rates of improvements. These technologies include the
recording density of magnetic disks (Figure 2.1.o) and tape (Figure 2.1.p), the bits per dollar
stored of magnetic tape (Figure 2.1.q), and the number of sequenced DNA base pairs. Since
these technologies have experienced accelerations after experiencing a slowing in the rates of
improvement, perhaps similar things will happen for organic transistors, computer
tomography, MRAM and FeRAM. Without more data, one cannot conclude that diminishing
returns will naturally lead to limits.
Focusing on the accelerations, previous analyses of S-curves have sometimes interpreted
them as the first half of an S-curve for a new technology26. Again, we urge caution in making
such an interpretation since the acceleration could be from a new technology whose
improvements in performance would appear as a straight line on a log plot. Since most of the
time series in Figure 2.1 appear as a straight line, this interpretation seems the more likely
than one of accelerations.
Nevertheless, we investigated the accelerations in more detail. To do so, we asked experts
about the cases of accelerations where the four of them (recording density of magnetic disks
and tape, bits per dollar stored of magnetic tape, and the number of sequenced DNA base
37
pairs) can be placed in two categories: DNA sequencing and magnetic storage. For the first
one, we sent mail to the National Human Genome Research Institute asking them for the
reasons behind two accelerations in the number of base pairs per dollar, which occurred in the
late 1990s and in the mid-2000s. For the second acceleration they wrote: “The slope of the
line became steeper because of the introduction of the 454 technology to the graph in one
quarter and then Illumina in another quarter. I have information for those two platforms, in
addition to the 3730 platform for some quarters. See the attached (two sheets).” According to
these two sheets, the number of base pairs per cost vs. time for these two new sequencers
display a straight line. For the first acceleration, they were unable to respond because they
claimed that the NHGRI was not formed until after the first acceleration.
For magnetic storage, we contacted Prof. Singh Bhatia, former head of research at IBM
Magnetic Storage. He claims that the acceleration in the late 1990s is from the introduction of
giant magneto resistance, whose discoverer received the 2007 Nobel Prize in Physics. Giant
magneto resistance is a quantum mechanical magneto-resistance effect that is observed in
thin-film structures composed of alternating ferromagnetic and non-magnetic conductive
layers. According to Prof. Bhatia, the recording density of disks based on GMR did not
accelerate after GMR’s introduction and instead their performance curves resemble a straight
line on a logarithmic plot. This provides further evidence that the shape of a performance vs.
time curve more closely resembles a straight line on a logarithmic plot than an S-curve.
We now consider the technologies that did not exhibit changes in the rates of improvement
during their time series. In particular, we consider the possibility that changes in technology
occurred and the performance vs. time curves for these sub-technologies resemble an S-
curve. This is also sometimes done by academics and practitioners. However, if there are not
changes in the rates of improvements, it is difficult to conclude that “sub” technologies within
the main technology did experience changes that might be interpreted as S-curves.
38
Furthermore, even if there are changes in the rates, we should interpret them as new
technologies whose performance curves resemble a straight line since most of the curves in
Figure 2.1 do resemble straight lines.
For example, consider computers. Although some claim that the second mini-computer’s
large percentage improvement over the first one suggests a case of “punctuated
equilibrium”27, Figures 1.l and 1.m suggest that this larger improvement is not truly “large” in
the overall history of computers. Computers experienced 15 orders of magnitude
improvements in computations per kw-hour and 12 orders of magnitude improvements in
computations per second and dollar over a 60 to 70 year time period.
Furthermore, when looking at Figures 1.l and 1.m, changes in the rates of improvement
cannot be seen during the mid-1960s when mini-computers were introduced, in the mid-
1970s when personal computers were introduced, in the 1980s when laptops were introduced,
or in the 1990s when personal digital assistants were introduced. Instead, the plots show
incremental improvements during this time period and no data point stands out. This is
consistent with an assessment by computer scientists and economists28 that these changes in
product architecture had a small impact on performance and most of the improvements in
Figures 1.l and 1.m can be attributed to improvements in electronic components such as
transistors and integrated circuits.
3. Discussion
This paper’s data base of 25 different technologies, 32 unique measures of performance,
and 575 individual data points show that the performance vs. time curves do not match the
predominant viewpoint of an S-curve. Instead, we find that the performance vs. time curve
for most technologies closely resembles a straight line on a logarithmic plot and thus has
constant rates of improvement over many years. The consistent results across many different
technologies and many different sources is particularly notable. Although one would expect
39
that many different sources of data would lead to many different curves, the shapes of the
curves are relatively consistent. Nevertheless, we do recognize that variations among
technologies will exist and thus a different data base might lead to different results.
One reason this paper’s results disagree with the predominant viewpoint of an S-shaped
performance vs. time curve is that we used a better methodology than do previous studies. In
particular, we borrowed more comprehensive time series from engineering and science
journals, which are biased towards rapid improvements, logarithmic plots, and sometimes
best laboratory results. Technologies with rapid rates of improvement may operate differently
from slower improving technologies that have received emphasis in previous studies29.
Focusing on technologies with rapid rates of improvements also caused us to emphasize
logarithmic plots, which have not been used in many previous studies. Finally, this paper’s
usage of data from engineering and science journals provided best industry numbers and
sometimes best laboratory results, which are different from previous studies that focused on
individual firms and on exploratory products30.
There are some advantages to using data that are only based on commercialized products
since commercialization is an important issue. However, using data only from
commercialized products, particularly if they are exploratory products31, will lead to gaps in
the data and thus the illusion that there are long periods of time without improvements,
whereas un-commercialized improvements were in fact achieved in R&D labs. When
products are exploratory and most of the improvements are therefore occurring in
laboratories, data on the improvements in the laboratories should be included in an analysis
of performance vs. time curves.
Perhaps most importantly, we believe that our results are consistent with how R&D works
and particularly with how R&D currently works. First, our results remind us that R&D is
extremely incremental and cumulative. Improvements build from past improvements and the
40
extensions to the knowledge base that these improvements bring. These improvements are
supported by the increases in R&D that occurs as demand rises in the form of greater sales
and production32. Thus, it should not be surprising that the rates of improvement are relatively
constant over many years as each new improvement enables further improvements. After all,
an exponential rate of improvement means that the percentage improvements each year are
constant.
Second, this paper’s results remind us that R&D is decentralized and it has become even
more decentralized over the last 50 years. As opposed to the vertically integrated world of the
mid-20th century that was being reported by Foster (1985)33 in which accelerations might be a
result of a few firms or funding agencies moving research funds from old to new
technologies, most R&D is now carried out in a world of “open innovation” where
technologies are bought and sold34. As part of this world of open innovation, most
researchers, including ones in universities, government labs, and even in corporations, are
expected to create their own research plans and to publish something new and different.
Funding for new technologies is also highly decentralized. This enables and requires
researchers to quickly move their efforts to new technologies long before the improvements
in an old technology have slowed.
Thus, while the performance vs. time curve for exploratory technologies or those of
individual firms may be more likely to experience an early acceleration35 or run into limits,
when we sum S-curves for individual organizations together to form an industry curve, or
when best laboratory results are included, the early accelerations or later limits will likely
disappear. Simply put, different firms, universities, and individuals will probably have
different patterns of investment, and organizational approaches, and strategic intents in R&D
and when these different patterns are combined into an industry pattern, straight lines on
logarithmic plots will likely emerge. Furthermore, increases in the amount of research
41
funding for universities from a growing number of funding sources adds to this phenomenon
of smoother rates of change.
What are the implications of these results for firms and governments? First, we should not
plan for and expect dramatic jumps in performance or ratios of performance to cost as many
do. Such expectations can lead to overinvestment and thus bubbles as in the Internet bubble
of 2000 or the current clean energy bubble. In clean energy, many have assumed that
demand-based subsidies will accelerate the rates of improvements batteries36 when the actual
improvements are quite constant with some more rapid than others. Instead, we should plan
for and expect incremental improvements to occur at fairly constant rates of improvement.
These constant rates of improvement allow us to predict to some extent the future levels of
performance and/or cost.
Second, if governments or firms would like to increase the rates of improvement or prevent
them from declining, perhaps they should invest more in basic R&D. This was the case for
magnetic tape and disks and for DNA sequencing machines in which four unique measures of
performance experienced accelerations after new forms of the technologies were introduced.
These new forms of the technologies enabled faster rates of improvement. This suggests that
firms and governments need patient long-term investments in new technologies in order to
achieve constant rapid rates of improvement over many years.
Third, we must be very careful about so-called limits, which make it more difficult for
firms to achieve faster rates of improvement37. Although we do not deny the existence of
theoretical physical limits, it appears that most of the analyzed technologies are far from their
limits. Coupled with the accelerations that were identified, it cannot be assumed that the
emergence of a slowdown or diminishing returns suggest limits are close. Of course, a lack of
limits in the data might reflect a lack of recent data. Thus, newer data might show such limits
or the absence of data might suggest that a limit has been reached and this is why newer data
42
points are not available.
Fourth, we believe that individual firms should use industry-level rather than firm-level
technology performance curves as the basis for formulating their corporate strategy for R&D,
organization reconfiguration, and alliances. Industry curves provide a much better
understanding of a technology’s rate of improvements and potential for further
improvements. Focusing on firm-level curves increase the chances that a firm will incorrectly
perceive jumps or limits and thus under invest or over invest in a new technology, or adopt
organizational structures, resource or capability positions that do not fit with the
environmental dynamism. Competitors that can accurately sense the industry-level
technology dynamism will develop a better fitted strategy and gain competitive advantage.
4. Appendix
We focused on three things in our data analysis and collection. First, we ensured in the
data collection that our time series included the start of commercial production since an
acceleration is purported to occur when commercial production occurs. For newer
technologies, this means ensuring that the newest data points cover the time period in which
commercial production has begun. For older technologies, this means ensuring that the oldest
data points include the start of commercial production.
Second, for analysis, we first used a visual inspection of data to detect accelerations or
limits that are consistent with an S-curve. To aid the reader, we include the raw data in a
meta-figure (See Figure 2.1) that contains small charts for the technologies shown in Table
2.1. Decelerations were interpreted as diminishing returns and if there was a dramatic
slowdown, this was interpreted as the technology reaching its limits. If we noted an
acceleration early in the time series, we interpreted this as the first half of an S-curve. If we
noted an acceleration later in the time series, we attempted to understand the reasons for the
acceleration by interviewing experts in the field. If the expert cited a new technology, we then
43
asked whether the performance vs. time curve for the new technology was a straight line on a
logarithmic plot or did the technology experience an acceleration early in its time series that
is consistent with the first half of an S-curve.
Third, we tested the time series data using several models. These models included a linear
model (1), a log model (2), and a log model with both time and time squared (3), as shown
here:
(1) y(t) = α + βt + ε
(2) Log y(t) = α + βt + ε
(3) Log y(t) = α + βt + γt2 + ε
where:
y = dependent variable involving performance, a ratio of performance to cost, or a ratio of
performance to physical attributes
t = time
α, β, γ = constants
ε = error term
For each technology and dimension of performance, we did regressions of these three
models in order to determine which model provides the best explanatory results. The
explanatory results were evaluated by looking at the R-squared and the p-value for each
model. For equation (3) we looked at the coefficient for time squared in addition to the p-
value. If it was negative and significant, we interpreted this as the emergence of diminishing
returns. If it was positive and significant, we interpreted this as a hyper-exponential curve. In
combination with the visual inspection, this hyper-exponential might be interpreted as the
beginning of an S-curve.
Table 2.2 summarizes the shapes of the performance curves vs. time for 32 different
measures of performance of 25 different technologies. It summarizes the R-squared values
44
and p-values for a linear model of performance as a function of time, a log model of
performance as a function of time, and a log model of performance as a function of both time
and time squared. The R-squared values suggest that a logarithmic model is more appropriate
than a linear model for all of the technologies except Li-ion batteries. Only Li-ion batteries,
magneto RAM, and cellulosic ethanol have similar or higher R-squared for a linear model
than for a logarithmic model.
Looking at the logarithmic model more closely, the first-order logarithmic model has
high R2 (> 0.9) and low p values (<0.001) for 26 of the 32 performance vs. time curves. There
are four R2-values that are between 0.7~0.9 and one R2-value at 0.54 (for Ferro-electric RAM)
and there is one p value at 0.014 (for Ferro-electric RAM). Therefore, the first-order
logarithmic model fits with the performance vs. time data with good statistical significance.
That suggests exponentially growing performance of technologies, and a constant percentage
rate of improvement each year. This is consistent with the estimated annual percentage rates
of improvement that are shown in Table 2.1 (right most column); these rates are equal to
(10β−1) where β is the constant coefficient of the first-order term of time in the log models.
Visualized on a logarithmic plot, the performance vs. time curves appear to be straight lines
(Figure 2.1), for which the slopes are β (See Table 2.1).
Several technologies have a higher R-squared for a logarithmic model that includes both
time and time squared than for a logarithmic model that includes just time and there is a
significant p-value for the time squared coefficient. The ones with a positive coefficient are
the number of memory bits per dollar for both DRAM and flash memory, data speeds for
wireless cellular, and recording density for magnetic disks. The ones with negative
coefficients are computer-tomography, aircraft, crystalline silicon solar cells, and amorphous
silicon solar cells.
45
46
Table 2.2 Regression Analysis of Performance and Cost vs. Time for Various Technologies
Technology Dimensions of
measure
Number of
Data Point
Linear Model Log. Model Log Model with Time2 Term
R-Sq. P-
Value
R-Sq. P-
Value
R-Sq. P-
Value
Sign of T2
Light Emitting
Diodes (LEDs)
Lumen/package, red 15 .29 .02 .98 <.0001 .98 .38 Positive
Lumen/package, white 7 .24 .15 .93 <.0001 .97 .13 Positive
Lumens per dollar, red 14 .58 <.0001 .92 <.0001 .91 .95 Positive
Si Solar Cells kwH per Dollar 47 .74 <.0001 .99 <.0001 .99 <.0001 Negative
Organic Cells Efficiency 9 .94 <.0001 .98 <.0001 .98 .15 Positive
Amorphous Si Efficiency 17 .83 <.0001 .56 .0003 .99 <.0001 Negative
Superconducting
Cables
Current x length-
YBaCuO
11 .77 .0002 .97 <.0001 .98 .07 Negative
47
Li-ion Batteries Energy per volume 11 .93 <.0001 .91 <.0001 .90 .59 Negative
Logic Chips Number of transistors
per chip
10 .66 .003 .98 <.0001 .98 .3 Negative
Microprocessors 21 .34 .003 .99 <.0001 .99 .2 Positive
Dynamic RAM 13 .36 .02 .997 <.0001 .997 .54 Positive
Bits per dollar 40 .36 <.0001 .99 <.0001 .99 .02 Positive
Flash Memory 15 .37 .001 .96 <.0001 .99 .001 Positive
Camera chips Pixels per dollar 31 .63 <.0001 .99 <.0001 .99 .04 Negative
Light sensitivity 13 .79 <.0001 .99 <.0001 .99 .99 Negative
1/Pixel size 13 .77 <.0001 .99 <.0001 .99 .6 Positive
Organic
Transistors
Mobility 9 .68 .004 .95 <.0001 .97 .03 Negative
48
Computers Instructions/time/price 73 .07 .01 .93 <.0001 .93 .015 Positive
Computations/kw-hr 66 .11 .004 .98 <.0001 .98 .78 Negative
Computer
Tomography
1/(scan time x
resolution)
13 .68 .0003 .74 <.0001 .92 .00005 Negative
Magnetic Disks Areal density 28 .19 .01 .95 <.0001 .97 .00002 Positive
Magnetic Tape Mbits per volume 14 .43 .005 .92 <.0001 .96 .003 Positive
Mbits per cost 14 .167 .08 .85 <.0001 .93 .002 Positive
Ferro-
electric RAM
Storage Capacity 9 .63 .007 .54 .014 .74 .11 Negative
Magneto RAM Storage Capacity 8 .86 .002 .82 .003 .84 .26 Negative
Phase Change
RAM
Storage Capacity 7 .44 .04 .9 .0002 .88 .71 Negative
Glass Fiber Distance/decibel loss 8 .77 .003 .91 <.0001 .99 .001 Negative
Last Mile Bits per second 8 .60 .02 .99 <.0001 .99 <.0001 Negative
49
Wireline
Wireless, 100
meters (cellular)
12 .11 .16 .86 <.0001 .91 .03 Positive
DNA Sequencer Base pairs per dollar 11 .19 .026 .85 <.0001 .96 <.0001 Positive
Cellulosic
Ethanol
Output per cost 11 .97 <.0001 .93 <.0001 .97 .57 Positive
Aircraft Number of Passengers
Times Speed
12 .81 <.0001 .97 <.0001 .99 .002 Negative
Figure 2.1. Performance or Performance-to Price Ratios for Technologies Listed in Tables 1 and 2
50
1965 1975 1985 1995 20050.01
0.1
1
10
100 b. Lumens per Dollar (Red LEDs) vs. Time
1950 1960 1970 1980 1990 2000 20100.001
0.01
0.1
1
10
c. KwHours per Dollar vs. Time for Crystalline Silicon Solar Cells
2002 2006 2010 201410
100
1000
e. Current (Amps) x Length (meters) vs. Time for YBaCuO Superconductor
1985 1990 1995 2000100
1000
d. Energy (Joules) Per Volume (cc) vs. Time for Li-ion Batteries
1980 1990 2000 2010 20200.001
0.01
0.1
1
10
h. Millions of Pixels per Dollar vs. Time for Camera Chips
1960 1970 1980 1990 2000 2010 20200.0001
0.01
1
100
10000
White
Red
a. Lumens per Package for LEDs vs. Time
1970 1980 1990 2000 20100.1
1
10
100
c. Efficiency of Amorphous Silicon and Organic Solar Cells vs. Time
AmorphousSilicon Organic
1950 1970 1990 20100.000001
0.0001
0.01
1
100
10000
Micro-processors
Dynamic RandomAccess Memory (DRAM)
MOSLogic
f. Millions of Bits per Chip vs. Time
1960 1970 1980 1990 2000 20100.0001
0.01
1
100
10000
Dynamic RandomAccessMemory
Flash Memory
g. Millions of Memory Bits/Dollar vs. Time
51
1985 1990 1995 2000 2005 20101
10
100
i. Light Sensitivity (mV/sq micron) vs. Time for Camera Chips
1985 1995 20050.001
0.01
0.1
1
j. 1/Pixel Size (sq micron) vs. Time for Camera Chips
19801985199019952000200520100.000001
0.0001
0.01
1
100
k. Mobility (cm2/Volt-sec) of Organic Transistors vs. TIme
1930 1950 1970 1990 20101E+01
1E+04
1E+07
1E+10
1E+13
1E+16
l. Computations/second/kwhour vs. Time
1940 1960 1980 20001E-12
1E-09
0.000001
0.001
1
1000m. Thousands of Computations/ Second/Dollar vs. Time
1970 1980 1990 2000 20100.00001
0.001
0.1
10
1000
100000
n. CT Scanner: 1/(Scan Time x Resolution) vs. Time
1950 1970 1990 20100.001
1
999.999999999999
999999.999999996
o. Millions of BIts per Sqare Inch vs. TIme for Magnetic Disks
19501960197019801990200020100.01
1
100
10000
1000000
p. Millions of Bits per Volume (cc) vs. Time for Magnetic Tape
1950 1970 1990 20100.0001
0.01
1
100
10000
q. Millions of BIts/Dollar vs. Time for Magnetic Tape
52
0.0001
0.01
1
100
Phase Change RAM (Random Access Memory)Ferro-electric RAM
Magnetic RAM
r.Storage Capacity (GB) per Memory Chip vs. TIme
196419661968197019721974197619781980198219840.001
0.01
0.1
1
10
s. Distance per Loss (km/decibel) vs. Time for Optical Fiber
19801985199019952000200520100.1
1
10
100
1000
10000
100000
t. Last Mile Bandwidth (1000s of bits/sec) vs. Time
1920 1930 1940 1950 1960 1970 1980100
1,000
10,000
100,000y. Aircraft Passenger Miles per Hour vs. Time
1980 1990 2000 20100.01
0.1
1
10
100
1000
u. Millions of Bits per Second vs. Time for Cellular Telecom
2000 2002 2004 2006 2008 2010 20120.1
1
10
x. Output (liters) per Dollar vs. Time for Cellulosic Ethanol
1970 1980 1990 2000 20100.0001
0.01
1
100
10000
1000000
100000000
v. Sequenced Base Pairs per Dollar vs. Time
53
Chapter 3
The Myth of Slowdowns in an Old Technology…..
A second myth about technology change is that slowdowns in old technologies drive the
development and also improvements in new technologies. This myth is partly derived from
the theory of S-curves for the performance vs. time of technologies (See Figure 1.1). The
idealized version for S-curves includes chains of S-curves in which a slowdown in an old
technology coincides with an acceleration in a new technology (See Figure 1.2). The
slowdown in the rate of improvement for an old technology is hypothesized to cause research
funds to move to a new technology and thus the new technology’s rate of improvement begins
to accelerate38.
There are many ways to test for whether slowdowns drive the development and
improvements in new technologies. Testing the idealized version of linked S-curves in which
a slowdown in an old technology coincides with an acceleration in a new technology is one
option. However, since Chapter 2’s analysis only found one technology that experienced an
acceleration in performance early in a technology’s life cycle, there will probably be few
technologies in which a slowdown coincides with an acceleration.
A less restrictive test is to statistically analyze whether rates of improvement for an old
technology are slower during a later than an earlier time period where these time periods
overlap with the introduction of a new technology (s). One key issue is the specific time
periods to test for slowdowns. One could test for a lag between a slowdown in an old
technology and the introduction of new technologies. Perhaps a slowdown in old technologies
cause new technologies to emerge and be improved a certain number of years later, say five or
ten years?
We simplify this problem by focusing on the first performance data that is recorded for the
new technology. This performance data could be for a laboratory result or for a commercially
introduced product as was discussed in Chapter 2. Building from the data collected for the
analysis in Chapter 2, time-series data from engineering and science journals (see Appendix
for more details) was analyzed for 15 pairs of old and new technologies. We tested for
whether the rates of improvement for the old technology are slower before or after a
performance metric was recorded for a new technology. We did this by testing whether the
slopes of the performance vs. time curves are statistically smaller before or after a
performance metric has been recorded for a new technology39.
54
We also reviewed the early history of these technologies in order to better understand the
motivation for the improvements and their timing. In particular, what caused the first or first
few performance metrics to be recorded? Was it a demand for a new technology or was it an
advance in science? Addressing these questions enables us to discuss these technologies
within a broader debate concerning demand-pull and technology-push. As noted in Chapter 1,
proponents of technology push argue that technology change is driven by universities and
other laboratories that “push” technologies into the marketplace while others argue that
technology change is driven by a market that “pulls” ideas for new technologies from the
minds of scientists and engineers in universities and laboratories through targeted research.
2. The Evidence
Table 3.1 summarizes rates of improvement for technologies that are potential substitutes
for each other in seven domains. The rates of improvements are for specific time periods and
specific performance metrics. Many of the measures include both cost and performance
factors in a single measure. Some of the rates are quite rapid, particularly those of more recent
technologies, while others are much slower, particularly those for technologies used in the 19 th
and early 20th centuries. For example, the rates of improvement for illumination by fire,
incandescent, and fluorescent lighting and for electricity from fossil fuels are much lower than
are the rates for the other technologies listed in Table 3.1.
Table 3.2 pairs a new technology(s) with an old technology(s) for the seven domains and it
presents the rates of improvement for the old technologies, both before and after performance
metrics are available for new technologies. While any new technology can be paired with any
of the previous technologies for which performance and/or cost data is available, Table 3.2
relies on the order in which these technologies have been developed and improved. Thus, it
pairs a new one with the previous one that was developed even though some of the ones
classified as old technologies are relatively recent ones and they have only started to be
implemented (e.g., LEDs).
Plots of the data for these old and new technologies are shown in Figures 3.1 to 3.9. There is
one figure for each domain with two for lighting and two for electricity. Each of the nine
figures includes time-series data for the old and new technologies in a single domain.
Although most of the old and new technologies are measured using the same dimension of
performance, some of the old and new technologies are measured using different measures of
performance. Quantum computers are measured by a different dimension than are digital
computers, electricity from nuclear is measured by a different dimension than are electricity
55
from fossil fuels and solar cells, and carbon nano-tube integrated circuits (ICs) is measured
from a different dimension from CMOS ICs. Nevertheless, from these individual figures one
can use visual inspection to judge if there is a slowdown in an old technology while a new
technology is being improved. As is discussed in the sub-sections below, none of the old
technologies in the smaller figures exhibit obvious slowdowns as new technologies are
improved.
Table 3.1. Rates of Improvements for Various Technologies, by Domain
Technolog
y Domain
Technology Dimension of Measure Time Period Rate Per
Year
Integrated Circuits
CMOS Microprocessor
1/Clock Period (Speed) 1986-2012 24%
Super-conducting Josephson Junction
1990-2010 20.3%
Carbon Nano-Tubes for Transistors
Purity of Carbon Nano-Tubes
1998-2012 52%
Non-Volatile Memory
Flash memory Storage capacity 1992-2013 50%
Resistive RAM 2006-2013 272%
Ferroelectric RAM 2001-2009 37.8% Magneto RAM 2002-2011 57.8%
Phase Change RAM 2004-2012 63.1%
Computing Digital Instructions per unit time 1947-2009 50%
Quantum Number of Qubits 2002-2012 107%
Telecom Last Mile Wireline Bits per second 1982-2010 48.7%
Cellular 1996-2013 79.1%
LAN 1995-2010 58.4%
56
WLAN 1996-2008 77.8%
Electricity Nuclear Dollars per kw of capacity 1950-1983 -2.7%
Fossil Fuel Dollars per Watt-Hours 1892-1972 5.5%
Crystalline Silicon Solar
1957-2003 15.9%
Efficiency 1957-2001 2.2%
Lighting Fire Luminosity per Watt 1855-1917 3.8%
Incandescent 1882-1948 2.2%
Fluorescent 1950-2002 0.5%
LEDs Lumens per package, red 1965-2008 16.8%
OLEDs Luminosity per Watt, green
1987-2005 29%
Displays LCDs Square meters per dollar 2001-2011 11.0%
OLEDs Luminosity per Watt, green
1987-2005 29%
Quantum Dots External Efficiency, red 1998-2009 36.0%
LEDs (Light-Emitting Diodes), OLEDs (Organic LEDs), LCDs (Liquid Crystal Displays), RAM (Random Access Memory). Sources: (Singularity.com 2014; Fujimaki, 2013; Franklin, 2013; ISSCC, 2013; Koomey et al, 2011; D-Wave, 2013; Brown, 2011; ISSCC, 2013; Hirsh, 1989; Nemet, 2006; Azevedo et al, 2009; Haitz and Tsao, 2011; Lee, 2005; Sheats et al, 1996; Economist, 2012, Kwak, 2010)
57
Table 3.2. Rates of Improvement before and after Improvements in New Technology Occur
Technology
Domain
New
Technology
Old
Technology
Rates of Improvement for Old Number of Data Points P-Value
Before New After new Before New After New
Integrated Circuits
Superconducting Josephson Junctions
CMOS micro-processor
1976-1990: 19.4% 1990-2007: 41% 15 17 <.001
Carbon Nano-Tubes for Transistors
1976-1998: 27.1% 1998-2007: 34.9% 23 9 .00314
Non-Volatile Memory
FeRAM NAND Flash memory
1992-2001: 51% 2001-2013: 50% 9 13 .39
MRAM 1992-2002: 45% 2002-2013: 55% 10 12 .20
PRAM 1992-2004: 53% 2004-2013: 47% 12 10 .86
RRAM 1992-2006: 48% 2006-2013; 55% 14 8 .69
Computing Quantum computers
Digital computers
1947-2002: 50% 2002-2009: 113% 55 13 .524
Telecom WLAN Wireline 1982-1994: 46% 1998-2006: 54% 4 4 .663
Cellular 1983-1995: 3.4% 1995-2012: 87% 3 10 <.001
58
LAN 1979-1995: 33.3% 1995-2004: 77.8% 3 3 .0029
Electricity Crystalline Silicon Solar
Fossil Fuel 1882-1957: 5.4% 1957-1972: 2.7% 14 4 <.001
Nuclear 1882-1947: 5.4% 1952-1972: 2.6% 12 5 <.001
Lighting Fluorescent Incandescent 1950-1970: 0.65% 1970-2002: 0.37% 10 2 .0012
OLEDs LEDs 1968-1986: 38.7% 1986-2008: 37.0% 5 15 .99
Displays Quantum Dots OLEDs 1987-1996: 35% 1996-2011: 22% 3 8 .90
59
Table 3.2 also summarizes a statistical analysis of the rates of improvement in Figures 3.1 to
Figure 3.9. The rates of improvement (i.e., slopes of lines on logarithmic plot) in the old
technologies were analyzed to test whether their rates of improvement slowed after performance
data were recorded for a new technology and to test whether any of the slowdowns are
statistically significant. The latter is done by analyzing whether the slopes of the lines, i.e., rates
of improvements, are statistically different for before and after performance data is recorded for
a new technology.
Table 3.2 shows that 7 of the 15 old technologies had slower rates of improvement after
performance data were recorded for a new technology than before performance data were
recorded. These old technologies are NAND flash memory, electricity from fossil fuels, two
forms of lighting (fluorescent and LEDs) and OLED displays. Two cases were found for flash
memory, once with respect to Ferro-electric RAM (random access memory) and Phase Change
RAM (PRAM), and two cases for electricity from fossil fuels, once with respect to nuclear and
once with respect to solar cells. Of these seven cases, the change in the slopes, i.e., rates of
improvement, were not large and only two were statistically significant at the 0.01 level.
Electricity from fossil fuels had statistically significant slower rates of improvement after the
first performance data were recorded for nuclear power and solar cells than before they were
recorded. The following sub-sections use this data to discuss all twelve of the new technologies
in more detail.
2.1 Integrated Circuits
Integrated circuits (ICs) have experienced very rapid improvements (about 35%) over the last
50 years and have become one of the most important technologies of the late 20 th and early 21st
centuries, enabling the creation of many new industries40. Their importance has also caused
many new technologies to be developed in order to maintain their rapid rates of improvement of
60
which these new technologies can be considered either complements or substitutes for ICs.
These new technologies include new structures for silicon-based transistors and memory cells,
three dimensional stacking of these transistors and memory cells, new semiconductor materials
such as gallium arsenide and gallium nitride, other types of materials such as carbon nano-tubes
and graphene, and more radical approaches such as atomic transistors.
The main question that this chapter asks is “are these new technologies being developed and
improved in response to a slowdown in the rate of improvement for the old technology,” in this
case silicon ICs such as microprocessor, memory, or other forms of ICs? Although the most
common measure of progress for ICs is the number of transistors per microprocessor or memory
bits per memory chip (typically called Moore’s Law), data on cost per transistor or memory cell
and the speed of microprocessor ICs are also available.
We have time series data for two new technologies, rapid single flux quantum (RSFQ)
superconducting Josephson junctions and carbon nano-tubes. Named for their discoverer,
Josephson junctions consist of a thin non-superconducting material that is sandwiched between
two superconducting materials and for which quantum tunneling can occur across the non-
superconducting material. These junctions can be used to construct various electronic devices of
which RSFQ devices are one of them.
Carbon nano-tubes (CNTs) are composed solely of carbon atoms, just as graphite, diamond,
and graphene are. While graphene is a one-atom thick layer of carbon atoms, one can think of
CNTs as graphene that is rolled into cylindrical tubes with either open or closed ends. These
CNTs can be produced with single, “few,” or “multi” walls of which the single wall ones with
high purity have the highest performance (e.g., conductivity) and cost. Suppliers of ICs are
interested in ultra-pure single wall CNTs because of their high conductivities, which contribute
to faster speeds in ICs.
Because speeds are important to existing and new ICs, we used the data on this measure for
61
silicon ICs - the clock period for microprocessor ICs. Figure 3.1 plots clock period data for
microprocessor ICs and superconducting Josephson junctions and purity data for carbon nano-
tubes. As noted in the previous paragraph, ultra-high purity is needed for high-performance
CNT-based ICs. Figure 3.1 and Table 3.2 show that the rate of improvement in the clock period
for microprocessor ICs was actually higher for the time period following the recording of
performance metrics for superconducting Josephson Junctions and CNTs than before the
recording of the performance metric.
This strongly suggests that a slowdown in the rate of improvement for microprocessor ICs did
not drive the development and improvement in either superconducting Josephson Junctions or
CNTs. Instead, the development and improvements in both superconducting Josephson
Junctions and CNTs were probably driven more by supply-side events than the rates of
improvements in the old technology. The first paper on RSFQ superconducting Josephson
junctions was published in 198541 a few years before the first recording of a performance metric
in 1990. CNTs were first fabricated in 1991 by Sumio Iijima42 and this was not long before the
performance metric for the purity of carbon nano-tubes was recorded in 1998. Furthermore,
improvements in complementary technologies such as processing and measurement capabilities
are also a reasonable possibility, perhaps borrowed from the semiconductor industry.
Figure 3.1 Old and New Integrated Circuit Technologies
62
1975 1985 1995 2005 20151.00E+06
1.00E+08
1.00E+10
1.00E+12 0.01
0.1
1
10
100
CMOS Microprocessors (Speed)
Speed (Hz) SuperconductingJosephsonJunctions (Speed)
CarbonNanoTubes(% impurities)
% Impurities
2.2 Non-Volatile Memory
Rates of progress for several other new IC-related technologies were also found and these
technologies are typically considered substitutes for one type of memory IC, called flash
memory. Flash memory is a type of memory that retains the memory values even when the
power is turned off and is therefore also called non-volatile memory. A key measure of progress
for memory ICs is the number of memory bits per chip and time series data for this metric were
found for flash memory and four types of new non-volatile random access memory (RAM): 1)
Ferro-electric RAM (FeRAM); 2) magnetic RAM (MRAM); 3) Phase Change (PRAM); an 4)
Resistive RAM (RRAM).
As shown in Table 3.1 and Figure 3.2, increases in the number of bits per chip have been
achieved over the last decade for all of these technologies. Figure 3.2 and Table 3.2 show that
the rate of improvement in the number of bits per chip are almost the same for all of the
technologies and the same for flash memory both before and after performance metrics were
recorded for the four forms of non-volatile memory. More specifically, the rates of improvement
for flash memory slowed slightly after performance metrics were recorded for FeRAM and
PRAM and increased slightly after performance metrics were recorded for MRAM and ReRAM.
63
Table 3.2 shows that none of these cases are statistically significant at the .01 or .1 levels.
This suggests that a slowdown in the rate of improvement for flash memory ICs did not drive
the development and improvement in new forms of non-volatile memory. Instead, the
development of and improvements in the new forms of non-volatile memory were probably
driven more by supply-side events than the rates of improvements in the old technology.
Improvements in FeRAM followed closely its invention by NASA’s Jet Propulsion Laboratory
in 1991 and subsequent developments by Ramtron, a fabless semiconductor company, and
Fujitsu. Improvements in MRAM followed the development of GMR (giant magneto resistance)
in 1988 and subsequent developments by Motorola from 1995. Improvements in PRAM
followed the development of new materials with fast crystallization speeds and those in ReRAM
followed the development of improved thin-film oxides43.
Figure 3.2 Old and New Non-Volatile Memory Technologies in Terms of Storage Capacity
1990 1995 2000 2005 2010 20150.0001
0.01
1
100
10000
NAND PRAM
FeRAM MRAM
RRAM
Billionsof Bits
Nevertheless, the state of current IC technology, whether this is microprocessor or memory
ICs, has probably had some effect on the development of new technologies for ICs. The possible
limitations for ICs have been discussed for decades and these discussions have probably
64
increased over the recent decades as some believe we are approaching the “limits” for ICs.
Photolithography was an early concern that has continued to be a challenge for the
semiconductor industry and this and other concerns are analyzed by the industry each year in the
International Technology Roadmap for Semiconductors44. Thus, although the initial development
of new memory and microprocessor technologies were probably not driven by a slowdown in
the current IC technology, concerns about the current state of the old technology have probably
motivated the continued improvements in these new technologies where these improvements are
supported by advances in science and developments in complementary technologies.
2.3 Digital and Quantum computers
One application for superconducting Josephson junctions is in quantum computers. Quantum
computers differ from conventional computers in that bits can be both 0 and 1 at the same time
according to a probability distribution. The bits in a quantum computer are called qubits and by
coupling multiple qubits, the performance of a quantum computer rises at a faster rate than do
increases in the number of qubits. While conventional computers operate on a base two system,
i.e., 0 or 1, and thus performance rises linearly with increases in the number of bits, the
performance of quantum computers rises non-linearly as the number of qubits are increased45.
Rates of progress for quantum computers were found with respect to number of qubits, qubit
lifetimes, and bits per Qubit lifetime; we focus on the number of qubits for simplicity although
the results are similar for the other two metrics since the first performance metrics for all of
them were recorded about the same year. Figure 3.3 shows the increases in the number of qubits
as well as the number of computations per kwH for digital computers. Figure 3.3 and Table 3.2
show that the rate of improvement for digital computers was actually faster after the recording of
the first performance metric for quantum computers in 2002, but the increase is not statistically
significant at the 0.01 level.
65
Figure 3.3 Old and New Computing Technologies
1940 1950 1960 1970 1980 1990 2000 2010 20201.00E+02
1.00E+04
1.00E+06
1.00E+08
1.00E+10
1.00E+12
1.00E+14
1.00E+16
1
10
100
1000C
ompu
tatio
ns p
er k
wH
Num
ber
of Q
ubits
QuantumComputers(number ofQubits)
Since a slowdown did not occur, this suggests that a slowdown did not drive the development
and improvements in quantum computers. Instead, like the other cases, the development and
improvements in quantum computers were probably driven more by supply-side considerations
than the rates of improvements in the old technology. The development of quantum computers
appear to coincide with the improvements in superconducting Josephson junctions that are
discussed in a previous sub-section. As these improvements have occurred, engineers and
scientists have used them to develop quantum computation, with D-Wave, a Canadian startup
being the most famous supplier of a quantum computer.
2.4 Telecommunications
Telecommunications has experienced rapid rates of improvements in both voice and data
communication. For data-related telecommunications, measures of progress include
transmission speeds and cost per transmitted data. For wireless data, transmission speeds per
band of frequency spectrum is also relevant since frequency spectrum is a limited resource. Old
and new technologies include several forms of copper cable, optical fiber, and wireless
66
technologies. Ideally, we would like to test whether a slowdown in the rate of improvement for
copper cable caused the development and improvements in optical fiber, or whether a slowdown
in wireline caused the development and improvement in wireless technologies.
We were able to find data on transmission speeds for four different technologies: 1) last mile
wireline; 2) cellular; 3) local area network (LAN); and 4) wireless LAN (WLAN). The first one
is not really a technology since a number of different copper cable and optical fiber technologies
have been used in the so-called “last mile” that connects individual residences to the Internet.
For example, the speed of copper cable for the last mile was increased through the use of faster
ICs and new techniques such as ADSL (Asymmetric digital subscriber line) and VDSL (very-
high speed digital subscriber line) before optical technologies were introduced46.
A second challenge for analyzing these technologies involves the time periods for which data
was obtained. Since the time periods for three of these technologies substantially overlap, it is
difficult to contrast most of the possible paired technologies. For example, since the data for the
last mile is only available since 1979 and thus few data points are available for the time period
before optical cable was introduced, it is difficult to test whether there was a slowdown before
optical cable was introduced (the improvements in the last mile that were mentioned above
suggest there was not a slowdown).
In any case, the only “new” technology that can be empirically compared with an old
technology is WLAN. Since WLAN might be considered a replacement for last mile wireline,
cellular, and LAN, we analyzed the rates of improvement for these three technologies both
before and after a performance metric was recorded for WLAN. As shown in Figure 3.4 and
Table 3.2, the rate of improvements for the old technologies were higher after a performance
metric was recorded for WLAN than before it was recorded. Thus, we cannot say that a
slowdown in neither LAN, last mile wireline, nor cellular caused the development and
improvements in WLAN.
67
Figure 3.4 Old and New Telecommunication Technologies in Terms of Data Speeds
1979 1984 1989 1994 1999 2004 2009 20140.01
1
100
10000
1000000
100000000
WLANLAN
Cellular
Last mile wireline
Instead, like the other technologies covered in this chapter, the development and
improvements in WLAN were probably driven more by supply-side considerations than the
rates of improvements in an old technology. It is widely recognized in the wireless field that
improvements in wireless technologies, including cellular and WLAN, are driven by
improvements in ICs. Indeed, the development and improvements in WLAN were largely
determined by the availability of high performing and low cost47.
2.5 Electricity
Most electricity has been and still is being produced from fossil fuels. Although some
electricity has been generated from hydro-electric sources, most electricity were generated from
coal, oil, or natural gas throughout the first 70 years of electricity beginning in 1880. Nuclear
Power was introduced in the early 1950s, solar cells were introduced for special applications
(e.g., satellites) in the late 1950s, and other technologies such as wind turbines were also
subsequently introduced.
68
We were able to find time series data on U.S. kwH per dollar from fossil fuels and solar cells
as well as data on kw per capital cost for Nuclear Power (See Table 3.1). As shown in Figure 3.5
and Table 3.1, the rate of improvement in the kwH per dollar from fossil fuels slowed after the
introduction of nuclear power and solar cells. Before the first cost data is available for nuclear
power in 1950 and solar cells in 1957, the rates of improvement in the kwH per dollar from
fossil fuels was 5.4% between 1882 and 1947 and also 5.4% between 1882 and 1957. From
1952, following the introduction of nuclear power, the rate of improvement fell to 2.6% and also
averaged 2.6% from 1957 until 1972. Furthermore, if newer data was analyzed, we would most
likely find increases in the cost of electricity from fossil fuels as environmental regulations were
implemented in the 1970s and subsequent decades; similar things occurred with nuclear power
(See Figure 3.5). This suggests that a slowdown in the rate of improvements for conventional
methods of electricity generation, primarily fossil fuel may have stimulated the development and
improvements in nuclear power and solar cells.
Figure 3.5 Old and New Electricity Generation Technologies
1890 1910 1930 1950 1970 1990 20100.001
0.01
0.1
1
10
100
0.0001
0.001
0.01
kwh/$ kw/$
kw/$ of Capacity for Nuclear Power(grey)
kwh/$ for Electricityfrom Fossil Fuels(red) and SolarCells (blue)
An alternative explanation is that the development and improvements in nuclear power and
69
solar cells were driven by supply side factors. The development of the atomic bomb occurred in
the early 1940s and this development and the application of nuclear reactors to submarines after
WWII probably contributed to performance improvements for electricity generation from
nuclear power48. Thus, the slowdown in the rate of improvement for electricity from fossil fuels
may have just coincided with the development of nuclear power.
For solar cells, a number of key developments also occurred just before the reductions in the
cost of electricity from them began in 1957. Although the photovoltaic effect had been
recognized in the 19th century, the modern junction semiconductor solar cell was first patented in
1946 and the first silicon solar cell was constructed by Calvin Fuller, Daryl Chapin, and Gerald
Pearson in 1954 at Bell Labs. Perhaps equally importantly, the launching of satellites by the
Soviet Union in 1957 and the U.S. in 1958 created a demand for solar cells that was very price
insensitive. Solar cells that produced electricity hundreds of times more expensive than from
fossil fuels were used in satellites and other remote locations in the 1960s and 1970s where it
was difficult and expensive to construct power lines49. Thus, the early improvements in solar
cells were primarily driven by factors that had little to do with conventional sources of
electricity long before solar cells were considered as a replacement for fossil fuels
Nevertheless, the recent increases in R&D funding for solar cells and other clean energy
technologies such as wind turbines along with demand-based subsidies for them is certainly
from dissatisfaction with fossil fuels. However, rather than a slowdown in a rate of
improvement, it is probably more a recognition that there are hidden costs from fossil fuels, in
the form of climate change, costs that continue to rise and are expected to rise further in the
future. This can be interpreted as a slowdown, a negative trend that is driving the development
of solar cells and other forms of clean energy and interestingly driving discussions about
reviving nuclear power.
For solar cells, it is interesting to dig a little deeper and analyze the development of new forms
70
of solar cells. The U.S. National Renewable Energy Laboratory has recorded best laboratory
efficiencies for 24 solar cell technologies50. Did a slowdown in crystalline silicon solar cells,
which are the oldest and still most widely used type of solar cell, lead to the development and
improvement of newer forms of solar cells? Since cost data for crystalline silicon solar cells 51
suggests that a slowdown has not yet occurred, we cannot conclude from this cost data that a
slowdown in the old type of solar cells has driven the development and improvements in a new
types of solar cells.
A second way to analyze the impact of a slowdown in old solar cells on the development and
improvement in new forms of solar cells is with best laboratory efficiencies. Did a slowdown in
the improvement of best laboratory efficiencies for crystalline silicon, lead to the development
and improvement of new forms of solar cells? One reason for doing such an analysis is that a
slowdown in best laboratory efficiency will occur before one in cost will occur since the highest
and lowest possible efficiencies are fixed and there are theoretical limits for maximum
efficiency in solar cells and other technologies. Nevertheless, it is plausible that some decision
makers might use a slowdown in the rate of improvement in efficiency as a reason to develop
new forms of solar cells.
To aid in this analysis, Figure 3.6 includes the data for the main types of solar cells. As shown
in Figure 3.6, a slowdown in the rate of improvement for single crystalline solar cells did not
begin until the early 1990s, long after performance metrics were recorded for other types of
solar cells. Both cadmium indium gallium selenide (CIGS) and Cadmium Telluride (CdTe)
experienced improvements before 1975 and amorphous silicon from about 1975. Thus, one
cannot conclude that the development and improvements in these types of solar cells were
driven by a slowdown in the rate of improvement for best laboratory efficiencies of crystalline
silicon solar cells.
71
Figure 3. 6 Old and New Solar Cell Technologies in Terms of Best Laboratory Efficiency
1975 1980 1985 1990 1995 2000 2005 2010 20150
5
10
15
20
25
OrganicPerovskiteQuantum DotsCdTeCIGSAmorphous Silicon
CrystallineSilicon Polysiilicon
One might make the argument that more recently developed solar cells - organic, Perovskite,
and quantum dot solar cells - were driven by the slowdown for crystalline silicon solar cells. Or
perhaps they were driven by a slowdown the rate of improvements for best laboratory CIGS,
CdTe, and amorphous silicon solar cells? Looking at Figure 3.6, it is hard to identify a
relationships between slowdowns in old solar cells and improvements in new solar cells or to
discern any type of pattern from the large number of solar cell technologies that are being
simultaneously being improved.
Instead, we believe there are better explanations for the recent improvements in organic,
Perovskite, and quantum dot solar cells. The first one was noted above; concerns about climate
change have led to increases in the funding of solar cells and thus probably the search for new
forms of solar cells. Second, the search for these solar cells and the relative success of this
search are driven by supply-side factors such as advances in science and technology that are also
seen in related applications. These related applications include organic transistors and displays
for organic materials and quantum dots where the Perovskite solar cells use combination of
organic and inorganic materials52. The emergence of these related applications and the ones for
72
solar cells suggest the supply-side factors, in combination with an increased interest in climate
change, have supported the improvements in organic, Perovskite, and quantum dot solar cells.
2.6 Lighting
Artificial lighting has been available for thousands of years in the form of fire (controlled
combustion in candles and other artifacts) and more recently in the form of oil and other lamps.
Modern day lighting began with incandescent lighting and electricity in the 1880s and has
continued with the introduction of fluorescent lighting in the 1940s and recent developments in
light-emitting diodes (LEDs) and organic LEDs (OLEDs).
We were able to find time series data for each of these forms of lighting, albeit the number of
data points for fire, fluorescent lighting, and perhaps incandescent lighting are insufficient for
careful analysis (See Figures 3.7 and 3.8. The much lower rates of improvement for them than
for LEDs and OLEDs (See Table 3.1), also make the definition of a slowdown somewhat
problematic. The insufficient number of data points means that we can only meaningfully test
the impact of a slowdown of LEDs on the development and improvement on OLEDs.
Nevertheless, we have included incandescent lighting for completeness even though there is
only two data points for the time period following the recording of a performance metric for
fluorescent lighting.
Figure 3.7 Old and New Lighting Technologies in Terms of Luminosity per Watt
73
1850 1875 1900 1925 1950 1975 2000 20250.01
0.1
1
10
100
Fire (Candles, Lamps, Gas Mantles)
Incandescent
Fluorescent
Luminosity per Watt
Table 3.2 shows the results of the statistical analysis. The rates of improvement for LEDs are
higher for the time period following the recording of a performance metric for OLEDs than for
before the recording of this performance metric. Although the rates of improvement for
incandescent lighting are slower following the recording of a performance metric for fluorescent
lighting, the difference is not statistically significant at the 0.01 level, as expected.
Figure 3.8 Old and New Lighting Technologies (continued) in Terms of Luminosity per Watt
1965 1975 1985 1995 20050.02
0.2
2
20
200
red LED green LEDblue LED white LEDgreen OLEDs yellow OLEDsblue OLEDs white OLEDs
74
These results suggest that slowdowns in the rates of improvements for old lighting
technologies did not lead to the development and improvements of new lighting technologies.
Like the other cases, it is appears more likely that the development and improvements in the new
technologies were more due to supply-side considerations than slowdowns in the old
technologies. Improvements in fluorescent lighting, LEDs, and OLEDs closely followed
developments in these technologies. The development of fluorescent lighting closely followed
successful experiments at GE by a renowned physicist Arthur Compton. Improvements in LEDs
and OLEDs closely followed the development of LED prototypes at GE in 1962 and OLED
prototypes at Eastman Kodak in 198753.
Nevertheless, like the other technologies covered in this chapter, interest in the technologies of
LEDs and OLEDs has increased as problems with the older technologies are more widely
perceived. As with fossil fuels, concerns about climate change have increased interest in new
forms of lighting technologies that have higher luminosity per Watt. Thus, the performance and
cost of the incandescent and fluorescent lighting, including the slow rate of improvement,
probably do currently have an impact on R&D spending and thus probably improvements in
LEDs and OLEDs.
2.7 Displays
Electronic displays have also experienced large changes in technologies and more recent ones
have experienced rapid rates of improvement. Cathode ray tubes have largely been replaced by
liquid crystal displays (LCDs) for desk-top computers and televisions and also by plasma
monitors for televisions. Touch-screen LCDs are used in smart phones and tablet computers and
still newer technologies such as 3D LCDs, OLEDs, quantum dots, holograms, and virtual retinal
displays are being developed for many applications.
We were able to find time series data on three of these technologies: 1) LCDs; 2) OLEDs; and
3) Quantum Dots (See Table 3.1). However, data for LCDs from the 1980s was not found thus
75
their rate of improvement for the time periods in which the first performance metrics were
recorded for OLEDs and quantum dot displays is unknown. Part of the problem is that LCDs did
not exist in the same form in the 1980s that they did in the 2000s and thus it is difficult compare
them along the dimension that they are currently being measured - cost per square meter. These
changes in form from super-twisted nematic to passive and active matrix LCDs enabled
improvements in quality measures such as resolution and frame rate but unfortunately such data
were not found. Nevertheless, the rapid rates of cost reduction over the last 15 years in cost per
square meter suggests that there probably was not a slowdown in the rate of improvement for
LCDs following the development of OLEDs.
Returning to the two technologies for which better data on rates of improvement exist, neither
OLEDs nor quantum dot displays have been implemented yet on a wide scale. OLEDs were first
used in mobile phones in 200154 but LCDs are still used on a much larger scale in mobile phones
than are LCDs. Quantum dots are beginning to be used in some LCD televisions55 as
improvements in luminosity per Watt occur.
Nevertheless, testing whether a slowdown in the rate of improvement in OLEDs has impacted
the development of quantum dots can be done with the data shown in Figure 3.9 and Table 3.2.
As shown there, the rate of improvement for OLEDs was higher after a performance metric was
recorded for quantum dots than before the performance metric was recorded. Thus, one cannot
say that a slowdown in the rate of improvement for OLEDs led to the development and
improvements of quantum dots. Instead, it is more likely that advances in quantum dot
technology56 that drove their development and thus improvements.
3. Discussion
This chapter has examined one widely held view about technology change: slowdowns in the
rates of improvement for old technologies leads to the development and improvement of new
technologies. According to this theory, slowdowns in the rates of improvement for old
76
technologies cause R&D resources to move to new technologies thus enabling improvements in
the new technology57. This viewpoint about slowdowns in old technologies leading to
improvements in new technologies is partly derived from the theory behind S-curves for the
performance vs. time of technologies. The slowdown in the rate of improvement for the old
technology hypothetically causes research funds to move to a still newer technology and thus the
newer technology’s rate of improvement begins to accelerate58.
Figure 3.9 Old and New Display Technologies in Terms of Luminosity Per Watt
1980 1985 1990 1995 2000 2005 2010 20150.1
1
10
100
1000
0.01
0.1
1
10
100Green OLEDYellow OLEDBlue OLEDRed QDOrange QDGreen QDBlue QD
Effici
ency
Lum
inos
ity P
err W
att
There are many ways we could have analyzed this viewpoint. We could have tested the most
extreme version of this viewpoint: the linked S-curve model shown in Chapter 1 (see Figure 1.2)
in which a slowdown in an old technology coincides with an acceleration in a new. We chose not
to test this however since the chances of finding statistically significant results are very low and
since it requires high densities of data points in the relevant time periods. Instead, we used a less
restrictive test for a slowdown in an old technology impacting on new technologies. In spite of
using a less restrictive test, however, our empirical analysis finds little support for this view.
Although 7 of the 15 old technologies had slower rates of improvement after performance data
77
were recorded for a new technology, none of these changes in rates of improvement (i.e., slopes)
were large and only two were statistically significant; the rate of improvement for electricity
from fossil fuels were slower following performance metrics being recorded for nuclear power
and solar cells. In addition, even in these two cases, supply-side reasons may as good or even a
better explanation for the timing of nuclear power and solar cells than does a slowdown in the
rate of improvement for electricity from fossil fuels.
The large number of technologies that are being simultaneously developed as replacements for
the existing technologies studied in this chapter also raises a more general question concerning
the theory behind the linked S-curves59 in Figure 1.2. Not only was this chapter not able to find
support for slowdowns in old technology leading to accelerations in new technologies, it found
that many new technologies are being simultaneously developed as replacements for the existing
technologies and not one technology as the “linked” S-curve theory suggests.
The theory of slowdowns in old technologies leading to the development and improvement of
new technologies has never attempted to address the issue of multiple new technologies. Where
do all the new technologies come from? Do they all appear because of a slowdown in an old
technology? If this is the case, they would emerge simultaneously and it is clear that they did
not. A better explanation is that these new technologies are the result of advances in science and
technology that are occurring independently of each other in a decentralized system of R&D60.
This chapter’s results are also mostly consistent with Rosenberg and Mowery’s61 view of
technology push in which technology change is driven by universities and other laboratories that
“push” technologies into the marketplace. As opposed to a market that “pulls” technologies into
existence, the technology push viewpoint suggests that the development and improvement of
new technologies depends more on laboratory and other supply-side developments than on the
market. As noted in Chapter 1, one of Rosenberg and Mowery’s persuasive arguments for
technology push was that few of the needs addressed by the important “innovations” of the 20 th
78
century had been recognized before the new technology was developed62.
However, from a larger perspective, we see both technology push and market pull operating in
the technologies that were studied for this chapter. Although the initial development and
improvements in the new technologies did not seem to depend on the state of the old technology,
as the new technologies were developed and improved, the state of the old technology became
more relevant to the improvements in the new technologies, as did market forces. For example,
the environmental problems with fossil fuels have increased interest in solar cells and even
nuclear power and indirectly increased interest in higher efficiency lighting technologies such as
LEDs and OLEDs. Similarly, increasing concerns about possible limits for Moore’s Law and ICs
have increased interest in new technologies such as carbon nano-tubes, superconducting
Josephson junctions, and new forms of non-volatile memory (NVM) such as ReRAM, FeRAM,
MRAM, and PRAM.
Part of this combination of technology push and market pull factors comes from the long
development time for these technologies. The time series data collected for this chapter suggest
that technologies undergo long periods of improvements before they are commercialized. For
example, OLEDs, quantum dots, non-volatile memories, carbon nano-tubes, and quantum
computers experienced rapid improvements for more than 10 years and sometimes longer than
20 years primarily in university laboratories without large amounts of commercial production.
The early years can clearly be characterized as technology push but as these technologies
become economically feasible, market pull becomes more important. As part of this market pull,
appropriate market segments must be identified by carefully contrasting the new and old
technologies along the dimensions of performance that are important to users. This is also part
of the disruptive technology’s story in which new markets play an important role in the success
of many new technologies63. All of the activities associated with finding new markets,
segmenting them, and modifying the new technology for these segments can certainly be
79
considered market pull.
Furthermore, some of this long development time involves not just finding the first market
segment, but moving the technology from niche markets to larger markets. The time series data
collected for this chapter suggest that some technologies undergo long periods of improvements
before they are considered for the markets that may end up being their largest. This is certainly
the case with solar cells and LEDs. Solar cells have experienced rapid improvements for more
than 50 years and LEDs more than 40 years. Yet it is only now that electric utilities are
beginning to implement solar cells and consumers are beginning to purchase LEDs as light
bulbs. This long time lag means that it is difficult to make a market pull argument early in the
technology’s development but over time, clearly market pull becomes the more important of the
two.
80
Appendix
As with Chapter 2, time series data was collected from leading scientific and engineering
journals, annual reports by reputable scientific organizations, general technology publications,
archival social science publications, general technology web sites, and technology-specific web
sites. As we collected performance vs. time data, we looked for technologies that might be
considered substitutes for each other. In order to increase the number of technologies that could
be analyzed, our criteria for substitutes was very broad. Furthermore, even when we could not
find performance vs. time data for new technologies, we retained their names in order to include
them in a broader discussion of the statistical data and the overall issues of whether slowdowns
in old technologies lead to improvements in new technologies and of demand-pull vs.
technology push.
Nevertheless, borrowing data from engineers and science journals enabled us to find time
series data for a large number of technologies and in some cases a large density of data points.
Some of the performance data are for best laboratory results, some of them are for
commercialized products, and some of them combine best laboratory results for early prototypes
with best products for after commercialization occurs. We do not attempt to distinguish between
these cases since we are primarily relying on the reliability of leading scientific and engineering
journals to collect meaningful data.
Once we have this performance vs. time data, there are a number of ways to analyze the
impact of old technologies on new technologies. The easiest method is visual. By plotting the
performance vs. time curves for the old and new technologies (we use logarithmic scales), we
can look for patterns that suggest an impact of the old technology on the new technology. The
idealized version of linked S-curves in which a slowdown in an old technology coincides with
an acceleration in a new technology would be a pattern that suggests a large impact of the old
technology on the new technology.
81
A second method is empirical, of which there are many approaches. Testing the idealized
version of linked S-curves in which a slowdown in an old technology coincides with an
acceleration in a new technology is one option. One can statistically test for whether slowdowns
occur in the old technology, accelerations occur in the new technology and whether these
slowdowns and accelerations coincide. Not only would this require a high density of data points
for both the old and new technologies, it is unlikely that support for this hypothesis could be
found.
We believe that a less restrictive test in which there is a higher chance of obtaining
statistically significant results can better add knowledge to this empirical uncertainty. We
statistically analyze whether the rates of improvement for the old technology are slower during a
later than an earlier time period where these time periods overlap with the introduction of new
technology (s). One key issue is the specific time periods to test for slowdowns. One could test
for a lag between a slowdown in an old technology and the introduction of new technologies.
Perhaps a slowdown in old technologies cause new technologies to emerge and be improved a
certain number of years later, say five or ten years?
We simplify this problem by focusing on the first performance data that is recorded for the
new technology. This performance data could be for a laboratory result or for a commercially
introduced product. Thus, we test whether the rates of improvement for the old technology are
slower before vs. after a performance metric has been recorded for a new technology. We do this
by testing whether the slopes of the performance vs. time curves are statistically smaller before
or after a performance metric has been recorded for a new technology. We do this by using a
technique that tests for the differences between slopes64.
Finally, we also reviewed the early history of these technologies in order to better understand
the motivation for the improvements and their timing. In particular, what caused the first or first
few performance metrics to be recorded? Was it a demand for a new technology or was it an
82
advance in science? In addition to general books and articles, we searched for terms such as
invention on Google in order to identify recent advances that may have enabled the recording of
the first performance metric(s).
83
Chapter 4
The Myth of Similar Rates of Improvement
A third myth is that all technologies have the same potential for rapid rates of improvement.
While few people would claim this myth as part of their belief system, few people recognize that
some technologies have faster rates of improvement than do other technologies. Although most
engineers and scientists know the rates of improvement for the technologies in their fields, few
know the rates of improvement across different fields. And among social scientists, few even
recognize that there are differences, much less know them; exceptions are those that address
general purpose technologies65. But outside of these few exceptions, many social scientists, even
those involved with technology from a managerial, economic, or sustainability point of view,
will argue against informing people of these differences for a variety of reasons that often
include the old refrain, ‘it’s not just about technology,” as if considering rates of improvement
prevents ones from considering other factors.
This is highly problematic for people who deal with many technologies or who must consider
multiple technological solutions to problems such as those of climate change, sustainability,
transportation, or smart cities. How can one devise solutions to climate change or help students
devise such solutions if one does not know rates of improvement for different technologies and
particularly if one doesn’t recognize there are differences? A lack of good data on rates of
improvement make it hard to navigate between different technologies and thus different possible
solutions66.
This chapter analyzes rates of improvement from several data bases. Section 2 analyzes annual
rates of improvement for 120 technologies and shows that the annual rates of improvement vary
from minus 10% to plus 42%, a very large range of annual improvements. It also shows that the
data contains multiple distributions and these multiple distributions provide insights into how
84
the improvements occur. Not only are there no technologies with rates between 15% and 25%,
the small literature on the economics of improvements suggests that the technologies with
annual rates of improvement of less than 15% can probably be further divided into multiple
distributions; this chapter divided them into two distributions, one for fixed formula chemicals
and one for other technologies. Analyzing these different distributions can help us better
understand why the rates of improvement are different.
Section 3 presents data on technologies with recent annual rates of improvement that are
greater than 25% per year. Although 25% per year is a much higher rate of improvement than
some important technologies such as steam engines and railways experienced in the 18 th and 19th
centuries, most scholars agree that rates of change in the late 20 th century have been much higher
than in previous centuries so we set a rather high rate of 25% per year as a cutoff for defining
rapid rates of improvement.
The fourth section examines several explanations for why some technologies experience
annual rates of improvement that are greater than 25% per year. These include greater: 1)
production; 2) demand; 3) R&D; and 4) opportunities for improvements. Consistent with
previous research that focused on inter-industry differences in R&D67, we conclude that greater
opportunities for improvements is the largest reason for the differences in rates of improvement.
2. The Distribution of Rates of Improvement
This section presents and analyzes data on rates of improvement. Data on rates of
improvement are calculated from performance vs. time curves that are presented elsewhere 68 and
publically available on a website (http://pcdb.santafe.edu/). We use all of the technologies that
are presented on this website. We calculate rates of improvement from the changes in
performance and/or cost between the first and last data points and from the number of years
separating the first and last data points.
Figure 4.1 plots the number of technologies as a function of annual rates of improvement in a
85
bar graph. The percentages along the x-axis are defined as plus or minus one percentage point
except for those with annual rates of improvement exceeding 42%. Thus, the 9 technologies
shown for a 10% annual rate of improvement have rates that are between 9.0 and 10.9%.
Excluding the ones with annual rates greater than 42%, the mean for the 120 technologies is
about 5% and the standard deviation is about ….
Most of the technologies have slow rates of improvement and two have negative rates. Sixty
five and 89% of them have annual rates of improvement that are less than 9% and 15%
respectively, and the rest have rates of greater than 25% per year. The gap between annual rates
of 15% and 25% suggest that multiple distributions exist within Figure 4.1 and these multiple
distributions can help us better understand why some technologies have more rapid rates of
improvement than do other technologies.
-10 -8 -6 -4 -2 0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40>420
5
10
15
20
25
FIgure 4.1 Number of Technologies by Annual Rates of Improvement
Annual Rates of Improvement
What types of multiple distributions might exist in Figure 4.1? In addition to the technologies
with annual rates less than 15% and greater than 25%, there also may be multiple distributions
within the main group of technologies that have rates lower than 15%. One hypothesis is that
86
chemicals may constitute a separate group of technologies since they have unique characteristics
that probably impact on their rates of improvement. Their fixed formulas limit the possible
changes in product design, they benefit from increases in the scale of production equipment
more than do other technologies, and the relationship between the capital cost of their
production facilities and their volumes is highly regular. The capital cost for these chemical
factories is a function of plant size to the nth power where n is typically between 0.6 and 0.769.
The hypothesized reason for this highly regular relationship between capital cost and plant
size is due to the impact of geometry on costs; the cost of pipes varies as a function of radius
while the output from pipes varies as a function of radius squared. Similarly, the costs of
reaction vessels vary as a function of surface area (radius squared) while the output of a reaction
vessel varies as a function of radius cubed70. This causes the capital costs of their production
facilities to fall in direct proportion to increases in production volumes.
We investigate the possibility that chemicals constitute a separate distribution by plotting the
number of chemical and non-chemical technologies vs. rate of improvement in Figures 4.2 and
4.3. Both of these figures show normal distributions with different means and standard
deviations. The chemical technologies have an average annual improvement rate of 6.3% while
the non-chemical technologies have an average annual improvement rate of 4.1%. For the
chemical technologies, the differences in annual rates among technologies in Figure 4.2 are
probably from differences in the role of material costs and from differences in timing. Since the
increases in production volume impact on capital costs and not materials cost (unless the
increased volumes impact on upstream activities), chemicals with higher materials costs will
experience slower rates of cost reduction. Differences in timing will result from the fact that the
benefits from increases in scale are probably larger during early years when the scale-up is first
started.
87
0 2 4 6 8 10 120
2
4
6
8
10
12
14
FIgure 4.2 Number of Chemical Technologies by Annual Rates of Improvement
Annual Rates of Improvement
For the non-chemical technologies, it is interesting that these technologies have lower annual
rates of improvement than do the chemical technologies in spite of having more opportunities
for product design changes than do the chemical technologies. Non-chemical technologies have
greater opportunities for improvements through changes in product design since they don’t have
fixed chemical formulas as chemicals do71 and thus one would expect more rapid rates of
improvement for them than for chemicals. Since they have slower rates of improvement than do
chemicals, there must be much fewer benefits from increasing the scale of production equipment
for non-chemicals than from chemicals, a conclusion that is consistent with previous research72.
88
-10 -8 -6 -4 -2 0 2 4 6 8 10 12 1402468
101214161820
Figure 4.3 Number of Non-Chemical Technologies by Annual Rates of Improvement
Annual Rates of Improvement
3. Technologies with Rapid Rates of Improvement
This section presents data on technologies with annual rates of improvement that are greater
than 25% per year. As with previous chapters, this data was collected from scientific and
engineering journals such as Nature, Science, Phys Status Solidi and IEEE, annual reports by
reputable scientific organizations such as the International Solid State Circuits Conference,
archival social science publications, and general technology and technology-specific web sites.
The data is shown in Table 4.1. To aid the reader, the technologies are placed into 8
categories. The first six categories are the transforming, storing, and transporting of energy and
information, and the last two are for the transformation and transportation of living organisms.
As expected many of the technologies in Table 4.1 can be defined as information technologies.
The distributions of these rates of improvement are shown in Figure 4.4. Figure 4.4 shows
the number of technologies with rates of improvement that are between the specific rates of
improvement on the x-axis. For example, there are 7 technologies with annual rates of
improvement between 25% and 35% and 6 technologies with annual rates of improvement
89
between 35% and 45%. The number of technologies declines as the rates of improvement
increase. The next section focuses on the reasons for rapid rates of improvement.
30 40 50 60 70 80 90 100 110 120 130 140 200 3000
2
4
6
8
10
Figure 4.4 Rapidly Improving Technologies: Number of Technologies Greater Than Rates on X-Axis
Annual Rates of Improvement
4. Why do some technologies have very rapid rates?
This section addresses four possible reasons for the more rapid rates of improvement for the
technologies in Table 4.1. They are greater production volumes, demand, R&D, or opportunities
for improvements.
4.1 Greater production volumes
Analyses of cost by the economics literature primarily focus on the factory floor and links
cost reductions with cumulative production; an exception is Peter Thompson’s paper in Journal
of Economic Perspectives73. In what has been termed learning by doing, costs fall as firms learn
90
to produce a single design in a single factory more efficiently and thus with lower costs. Workers
become better at tasks and firms introduce better work flows, better process control, and
automated manufacturing equipment, and they promote organizational learning74. Thus, some
technologies might experience faster rates of improvement merely because they have a faster
growth in production volumes.
We give three reasons why greater production volumes are not the main reason for the rapid
rates of improvement in the technologies shown in Table 4.1 in addition to the reasons stated by
Peter Thompson. First, a previous section contrasted the rates of improvement for chemicals and
non-chemicals, both within the technologies experiencing annual rates less than 15%. That
section showed that chemicals have more rapid rates of annual improvements than do non-
chemical technologies in spite of the fewer opportunities for product design changes in
chemicals than in non-chemicals. That analysis and previous research75 suggests that this is
probably because chemicals benefit more from increases in the scale of production equipment,
which is one form of learning by doing and it requires increases in production volumes.
Nevertheless, since chemicals do not experience rates of improvement that approach those of the
technologies shown in Table 4.1, it is unlikely that rapid rates of improvement in Table 4.1 are
due to greater production volumes.
A second and perhaps more important reason why greater production volumes don’t explain
the rapid rates of improvement in Table 4,1 is that many of the dimensions are performance as
opposed to cost-related measures or they include both performance and cost in a single
dimension. Improvements in performance probably require product design changes76 and the
increases in production don’t impact on performance. We pursue this argument further in
Chapters 5 and 6. On the other hand, the increases in production volume might indirectly lead
to improvements in performance since increases in production is associated with increases in
demand and increases in demand often lead to increases in R&D spending. The next section
91
addresses demand.
4.2 Greater Demand
A second possible reason for faster rates of improvement is greater demand. We note that this
is actually a partial variant of learning curves since demand is related to cumulative production.
The difference between the impact of demand and production on costs is that increasing demand
can induce greater R&D and thus lead to lower costs and higher performance through the greater
R&D; Jacob Schmookler empirically demonstrated this “demand-pull” hypothesis in 196677.
This caused some scholars to consider cumulative production as a general proxy for effort and
thus the driver of new product and process designs and improvements in performance and cost78.
A major reason for why demand is an inadequate explanation for the differences in rates of
improvement were provided by Nathan Rosenberg and David Mowery in the 1970s concerning
demand-pull vs. technology push. Chapter 3 summarized some of arguments and their
agreement with the data that was presented in Chapter 3. Only a small number of the old
technologies analyzed in Chapter 3 exhibited a slowdown in their rates of improvement as a new
technology was improved thus suggesting that the data is more consistent with a technology-
push than demand-pull theory.
However, even before his work with David Mowery, Nathan Rosenberg had adequately
rebutted Jacob Schmookler’s over-emphasis on demand with a 1974 paper in the Economic
Journal. His response to Schmookler was that “some demand induces the inventions that satisfy
it. But which and when?” He then described many areas in which demand has existed for many
years if not centuries (e.g., health care), but was not satisfied until recently. He concludes his
paper with: “the economic question is: Given the state of the sciences, at what cost can a
technological end be attained?79” We offer a possible answer to Rosenberg’s question in sub-
section 4.3 and subsequent chapters.
92
4.3 Greater R&D Intensity
A third possible reason for the more rapid rates of improvement for the technologies in Table
4.1 is greater R&D intensity. Perhaps more R&D has been devoted to these technologies than
other technologies and this has caused the technologies in Table 4.1 to have more rapid rates of
improvement than do other technologies. Furthermore, this greater R&D might come from
greater production, as mentioned in the previous sub-section.
Research on inter-industry differences in R&D by Alvin Klevorick and his colleagues80
suggest that this is not an adequate explanation for the rapid rates of improvement in Table 4.1.
Research on inter-industry differences in R&D focus on two issues - differences in
appropriability and differences in opportunities – of which the second reason is generally
perceived to be the more important reason. Although differences in appropriability do impact on
firm R&D, the sign is not clear. While strong appropriability increases incentives to engage in
R&D, weak appropriability lowers the cost of research for others and thus differences in
appropriability does not probably explain the rapid rates of improvement in Table 4.1.
Differences in opportunities probably explains these differences in inter-industry R&D, a
conclusion that was reached in the early 1970s by Nathan Rosenberg81. Alvin Klevorick and his
colleagues based their conclusions on a survey of R&D managers. The survey asked managers
to assess the impact of various fields of science on technological progress and about
technological advances originating outside their industry. From this survey Klevorick and his
colleagues concluded that differences in opportunities best explain the differences in R&D
intensity among firms. Their conclusions is consistent with this paper’s analysis. The one
difference between this book’s conclusions and Kleverick’s conclusions is that there other ways
of measuring differences in opportunities for R&D in addition to the method used by Klevorick
and his colleague. The next sub-section and subsequent chapters describe a different approach to
explaining differences in opportunities than with surveys of R&D managers.
93
4.4 Greater opportunities for improvements
Building from the previous sub-section, a fourth possible reason for the rapid rates of
improvement in Table 4.1 is greater opportunities for improvements through greater product and
process R&D. R&D leads to more improvements per R&D dollar for some technologies than for
other technologies and these improvements lead to increased consumer welfare, profits, and
R&D spending. Thus, the technologies that are easier to improve end up having faster rates of
improvement where greater R&D on these technologies magnifies the differences in rates of
improvement among different technologies. This can lead to the dramatic differences between
technologies that are shown in Figure 4.1.
This argument is a variant of the argument made by Alvin Klevorick and his colleagues. They
explain different levels of opportunities in terms of the applicability of science to technological
progress and the technological advances originating outside their industry and they measure
these factors through surveys. This chapter and subsequent chapters focus on rates of
improvement where the applicability of science or technologies advances originating outside the
industry might impact on the rates of improvement. Thus, it could be that the factors suggested
by Alvin and Klevorick and his colleagues might be the ultimate explanation for differences in
rates of improvement but subsequent chapters provide a more detailed analysis of the product
and process design changes that enable the improvements and that may be supported by the
factors suggested by Alvin Klevorick and his colleagues.
5. Discussion
This chapter showed that annual rates of improvement range dramatically from minus 10% to
greater than 42%, a very large range. Two-thirds and 89% of the technologies had annual rates
of improvement that are less than 9% and 15% respectively and that there are multiple
distributions in this data base. In addition to the two distributions suggested by the gap between
94
annual rates of 15% and 25%, this chapter’s analysis also suggests that the technologies with
annual rates of improvement below 15% can be further divided into two distributions, one for
chemical and one for non-chemical technologies.
The chapter then analyzed possible reasons for the rapid rates of improvement, defined as
greater than 25% per year. To do this, data was gathered and presented for technologies with
these rapid rates and argued that these technologies are potential candidates for general purpose
technologies. It then discussed possible reasons for the rapid rates of improvement, which
included greater production volumes, demand, R&D intensities, and opportunities for
improvements. This discussion concluded that greater opportunities for improvements from
product and process designs are the probable reason for the rapid rates of improvement for some
technologies.
The next two chapters examine these product and process designs by demonstrating the
existence of two other myths. Chapter 5 examines the myth that product design changes drive
performance increases and process design changes drives cost reductions, with product
preceding process design changes in a technology’s life cycle. In doing so Chapter 5 shows that
process design changes by themselves cannot explain rapid rates of improvement in costs.
Chapter 6 examines the myth that costs fall as cumulative production rises in a learning curve. In
doing so, Chapter 6 analyzes 13 technologies that experienced rapid rates of improvement with
zero commercial production and identifies the product and process design changes that led to
these rapid rates of improvement.
95
Table 4.1 Technologies with Recent Rapid Rates of Improvement
Technology
Domain
Sub-Technology Dimensions of measure Time
Period
Rate Per
Year
Energy
Trans-
formation
Light Emitting Diodes
(LEDs)
Lumens per Dollar, white 2000-2010 40.5%
Organic LEDs Luminosity per Watt, green 1987-2005 29%
GaAs Lasers Power density 1987-2007 30%
Cost per Watt 1987-2007 31%
Quantum Dot
Displays, Solar Cells
Efficiency, red displays 1998-2009 36.0%
Efficiency of Solar Cells 2010-2013 42.1%
Energy
Trans-
mission
Superconductors:
BSSCO and YBCO
Current-length per cost 2004-2010 115%
Current x length - BSSCO 1987-2008 32.5%
Current x length - YBCO 2002-2011 53.3%
Information
Trans-
formation
Microprocessor ICs Number of transistors/chip 1971-2011 38%
Camera chips Pixels per dollar 1983-2013 48.7%
MEMS: Artificial Eye Number of Electrodes 2002-2013 45.6%
96
MEMS: inkjet printers Drops per second 1985-2009 61%
Organic Transistors Mobility 1982-2006 109%
Single Walled Carbon
Nanotube Transistors
1/Purity (% metallic) 1999-2011 32.1%
Density 2006-2011 357%
Superconducting
Josephson Junctions
Qubit Lifetimes 1999-2012 142%
Bits per Qubit lifetime 2005-2013 137%
Photonics Data Capacity per Chip 1983-2011 39.0%
97
Table 4.1 Technologies with Recent Rapid Rates of Improvement (continued)
Technology
Domain
Sub-Technology Dimensions of measure Time
Period
Rate Per
Year
Information
Trans-
formation
Computers Instructions per unit time 1979-2009 35.9%
Instructions per time-cost 1979-2009 52.2%
Quantum Computers Number of Qubits 2002-2012 107%
Information
Storage
Flash Memory Storage Capacity 2001-2013 46.8%
Resistive RAM Storage Capacity 2006-2013 272%
Ferroelectric RAM Storage Capacity 2001-2009 37.8%
Magneto RAM Storage Capacity 2002-2011 57.8%
Phase Change RAMStorage Capacity
2004-2012 63.1%
Magnetic Storage Recording density of disks 1991-2011 55.7%
Recording density of tape 1993-2011 32.1%
Cost per bit of disks 1956-2007 32.7%
Information
Transmission
Last Mile Wireline Bits per second 1982-2010 48.7%
Wireless, Cellular Bits per second 1996-2013 79.1%
98
Wireless, WLAN 1995-2010 58.4%
Wireless, 1 meter 1996-2008 77.8%
Biological
Trans-
formation
DNA Sequencing per unit cost 2001-2013 146%
Synthesizing per unit cost 2002-2010 84.3%
(Azevedo et al, 2009; Haitz and Tsao, 2011; Lee, 2005; Sheats et al, 1996; Martinson, 2007;
Kwak, 2010; NREL, 2013.Shiohara et al, 2013; Selvamanickam V 2011; Wikipedia, 2014; Preil,
2012; Chader, 2009; Stasiak et al, 2009; Hasegawa and Takeya, 2009; Franklin, 2013; Devoret
and Schoeldopf, 2013; Evans et al, 2011; Koomey et al, 2011; D-Wave, 2013; ISSCC, 2013;
Francis, 2011; Yoon, 2010; Brown, 2011; ISSCC, 2013; NHGRI, 2013; SingularityHub.com,
2013)
Chapter 5
The Myth of Product Design Changes Leading to Improvements in …..
A fourth myth is that product design changes lead to increases in performance early in the life
cycle and process design changes lead to reductions in costs later in the life cycle (See Figure
5.1). Like the S-curve, this myth is taught in almost every technology management course in the
world and it causes firm strategies and government policies to focus on product design changes
to achieve performance increases and to focus on process design changes to achieve cost
reductions82. But also like the S-curve, empirical analyses are few and these analyses have used
patents as opposed to actual performance and cost data to test the myth and found it to be
unsupported83.
99
This chapter does the first comprehensive analysis of cost and performance data to test this
myth. It analyzes the cost and performance of 17 different technologies, 22 unique time-series
pairs of performance and cost, 358 unique data points and 705 total years of data for these time-
series pairs. Technologies such as chemicals, agricultural products and materials are excluded
since they often have a fixed chemical composition and thus improvements in performance do
not occur and most of their improvements in cost are driven by process innovations or by
increasing the scale of their production equipment84. But for other technologies, the analysis
shows that improvements in performance and cost for most products are highly correlated over
many decades.
Such a correlation should not exist if product design changes drive performance increases and
process design changes drive cost reductions and that product precede process design changes.
Thus, these results suggest that new explanations for how improvements in performance and
100
cost occur are needed and several explanations are offered following the presentation of the
statistical analysis.
5.1 The Evidence
Table 5.1 summarizes the rates of improvement that were found in the engineering and
science literatures for both relative performance and relative price in 17 unique technologies.
Relative performance is a ratio of performance to mass, volume, area, chip, package, or best
possible output (i.e., efficiency) except for one dimension of performance for microprocessors.
Relative price is a ratio of performance to price since performance must be in the numerator in
order that both time series are moving in the same direction in order to calculate correlation
coefficients. To aid the reader, the technologies are placed into 3 general categories of which 9
can be roughly classified as energy, 7 as information, and 1 as biology. These general categories
are further divided into 7 sub-categories to help the reader understand the data set; these sub-
categories include transformation of information, energy, and biology, storage of information
and energy, and transport of information.
All of the technologies shown in Table 5.1 are for a single technology concept except for
TansAtlantic Cable, DNA Sequencing, and Magnetic storage (both tape and disks).
TransAtlantic Cable is a telecommunication cable that is placed under the Atlantic Ocean and it
has been done with both copper and fiber optic cable. DNA sequencing equipment has also
involved changes in the concepts that form the basis for them. All of the other technologies are
for a single concept.
Table 5.1. Annual Rates of Improvement for Specific Technologies
Technology
Domain
Sub-
Technology
Dimensions of measure Time
Period
Rate Per
Year
Infor- Microprocessor Number of transistors per chip 1971-2011 40%
101
mation
Trans-
formation
Integrated
Circuits (ICs)
Microprocessor clock speed 1976-1999 32%
Transistors per cycle time per price 1976-1999 89%
Memory
(DRAM) ICs
Number of transistors per chip 1971-2010 48%
Price per memory bit 1971-2000 43%
Flash Memory Memory bits per price 1992-2007 66%
Storage capacity 1993-2008 58%
Computers Computations per kwh 1946-2009 52%
Computations per time and dollar 1945-2008 38%
Infor-
mation
Storage
Magnetic Tape Bits per unit cost 1994-2011 33%
Bits per unit area 1994-2011 34%
Magnetic Disk Bits per unit cost 1956-2007 39%
Bits per unit area 1956-2007 43%
Information
Transport
TransAtlantic
Cable
Bits per second 1951-2001 36%
Bits per dollars per distance 1951-2001 25%
Living Biological DNA Sequencing, base pairs/cost 1971-2011 35%
102
Organisms transformation DNA Sequencing, bases/person/day 1986-2011 79%
Energy
Trans-
Formation
ICE Passen-
ger Car
Power per kg 1896-1994 5.4%
Power per liter 1931-1994 1.4%
Power per cost 1896-1994 6.2%
Aircraft Power per kg 1919-1945 7.7%
Power per liter 1919-1945 13.3%
Power per cost 1919-1945 7.6%
Light Emitting
Diodes (LEDs)
Lumen per package, red 1968-2005 38%
Cost per lumen, red 1973-2004 29%
Lumen per package, white 2000-2009 75%
Cost per lumen, white 2000-2009 39%
GaAs Lasers Power density 1987-2007 30%
Output per price 1987-2007 30%
Photo-sensors
(Camera chips)
Pixels per dollar 1983-2013 49%
Pixels per area (Resolution) 1987-2009 21%
Light sensitivity 1986-2008 18%
Silicon Solar
Cells
Power output per unit cost 1957-2001 16%
Efficiency 1957-2001 2.2%
103
Energy
Storage
Lead Acid
Batteries
Energy stored per unit volume 1882-2005 4%
Energy stored per unit mass 1882-2005 4%
Energy stored per unit cost 1950-2002 3.6%
Capacitors Energy stored per unit mass 1962-2004 17%
Energy stored per unit cost 1945-2004 4%
DRAM: Dynamic Random Access Memory; Magnetic Resonant Imaging; ICE: internal
combustion engine
Sources, from top to bottom:ICs (Singularity.com 2014; ICKnowledge, 2009); Flash Memory
(EDN, 2014; iNEMI, 2008); Computers (Koh and Magee, 2006; Koomey et al, 2009), Magnetic
Tape (Koh and Magee, 2006, Mellor, 2012, Coughlin, 2012); Magnetic Disk (Francis, 2011;
Yoon, 2010), TransAtlantic Cable (Koh and Magee, 2006); DNA Sequencing
(SingularityHub.com, Carlson, 2013); ICE (Koh and Magee, 2008): LEDs (Haitz and Tsao,
2110); GaAs lasers (Martinson, 2007); camera chips (Suzuki, 2010; futurefab, 2013); Solar
(Nemet, 2006; NREL, 2013); energy storage (Koh and Magee, 2008); superconductors
(Shiohara et al, 2013; Selvamanickam V 2011);
Although some might break down computers into different architectures such as mainframe,
mini, and personal computers, computer scientists claim they are all based on the same
underlying concept85.
Data was found for multiple measures of performance in five of the technologies. This
includes power per mass and per volume for passenger car and aircraft engines, pixels per area
and sensitivity for photo-sensors (energy transformation), energy stored per mass and per
volume for batteries and capacitors (energy storage), and number of transistors and clock speed
for microprocessors (information transformation). With 17 unique technologies and multiple
measures of performance for five of them, this gives us 39 different time series in Table 5.1 and
104
potentially 22 paired time series of relative or absolute performance and performance-to price
ratios.
Most of the technologies in Table 5.1 were experiencing rapid improvements in cost and
performance and thus the increases in performance-to price ratios probably reflect reductions in
cost and not just price. For the performance-to price ratios, 11 of the 17 technologies had rates of
improvement of greater than 30%, 2 had rates between 20% and 29% and only three had less
than 10%. These three are internal combustion engines for passenger cars and aircraft,
capacitors, and lead acid batteries. The rapid rates of improvement suggest that the price
changes probably reflect changes in cost for most of the technologies. When the rates of
improvement are greater than 10%, 20% and 30%, performance-to price ratios will rise by 2.6,
6.2, and 13.8 times over ten years, reductions in cost will be much larger than will any changes
in margins, and thus prices and costs closely follow each other for these rapidly improving
technologies. Furthermore, since drops in profit margins often occur later rather than earlier in a
technology’s lifecycle, any drop in margins would cause an acceleration in the performance-to
price ratio during the latter part of the technology’s life cycle and thus reduce the chances of a
correlation between the improvements in cost and performance.
Figure 5.2 (See Appendix) is a meta-figure that summarizes in 22 smaller figures the time
series of relative performance and performance-to price ratios for the 17 technologies. These
figures are labeled from “a” to “v.” For each of the smaller figures, the left axes are for relative
performance and the right axes are for performance-to price ratios. The units for both of them
are shown in the title of each smaller figure. The axes are not labeled due to a shortage of space
in Figure 5.2.
These small figures show that relative performance and performance-to price ratios rose
steadily over time for most of the technologies. This is particularly true for all of the information
technologies, DNA sequencing, and most of the energy technologies. Only two of the energy
105
technologies show some ups and downs. The Watts per Dollar for Aircraft engines (internal
combustion) fell at the end of the time series while the Watts per kilogram (kg) and Watts per
liter rose. Second, the relative performance (Watts per kg and Watts per liter) of passenger car
engines rose and fell at the same time as did Watts per Dollar.
Table 5.2 summarizes the statistical analysis for the 22 different paired time series. It shows
that there were 358 unique data points in these paired time series or an average of 16 data points
per time series. The lowest was 6 data points and the highest was 55. It also shows that all of the
technologies covered 705 years in total or on the average 32 years per paired time series. The
lowest was 10 years (white LEDs) and the highest was 99 years (passenger car engines).
Since the lowest number was larger than the number of years that passed before a change
occurred in industry emphasis for personal computers from performance to cost (and the
emergence of a dominant design)86, the data set can be used to determine whether relative
performance is important early in the life cycle and relative price becomes more important later
in the life cycle, which would not be the case if there is a correlation between the time series
data for relative performance and relative price. As discussed in the literature review, if
increases in performance are driven by changes in product design and reductions in cost are
driven by changes in process design and if the product design changes precede those of process
design changes, we would not expect improvements in performance and cost to be correlated.
Table 5.2 summarizes the correlation coefficients for the 22 different paired time series. The
coefficients exceed 0.9 for 17 of the 22 unique measures. Only two of them fall below 0.5 and
these are for internal combustion engines in aircraft. The other three measures of performance
fall between 0.5 and 0.9 with two of them at 0.895. Looking more closely at Table 5.2, 1 of the
22 measures of performance are the same for both the “relative performance” and for the
performance in the performance-to price ratios. This is the speed of microprocessors. As might
be expected, there is a high correlation between improvements in relative performance and
106
improvements in the performance-to price ratios since the measures of performance are the
same. However, at 0.937, it is not much higher than the other correlation coefficients.
Furthermore, even if this technology is dropped, the other 16 technologies and 21 unique paired
time series show high correlations.
Overall, the results show that improvements in performance and improvements in cost/price
are highly correlated over time periods that are very long (average of 30 years) and thus the
hypothesized change from performance to cost and from product to process design that are
emphasized in the literature should have occurred during these long time periods. However, the
strong correlations suggest that the changes did not occur and the long time periods suggest that
they are unlikely to occur in the future for most of the technologies.
Table 5.2. Correlation Analysis of Performance and Cost over Time for Various Technologies
Technology Dimensions of measure Time
Period
Data
Points
Correlation
Micro-
processor ICs
Number of transistors per chip vs. number of
transistors/cycle/$
1976-2011 19 .937
Speed (Hz) vs. number of transistors/Hz/$ 1976-2007 18 .947
Memory (DRAM)
ICs
Number of transistors per chip vs. price per
memory bit
1971-2000 13 .993
Flash Memory Memory capacity vs. bits per price 1992-2008 10 .982
Computers Computations per kwh and computations per
unit time and price
1946-2004 31 .920
Magnetic Tape Bits per unit area and bits per unit price 1994-2011 13 .919
107
Magnetic Disk Bits per unit area and bits per unit price 1956-2007 11 .967
Trans Atlantic
Cable
Bits per second and bits per cost per distance 1951-2001 12 .999
DNA Sequencing Speed (bases per person per day) and base
pairs per cost
1985-2011 8 .999
ICE Passenger
Car
Power per kg and power per cost 1896-1994 55 .931
Power per liter and power per cost 1931-1994 45 .721
Aircraft Power per kg and power per cost 1919-1945 13 .487
Power per liter and power per cost 1919-1945 13 .343
Power per
LEDs (Red) Lumen per package and cost per lumen 1973-2005 11 .944
LEDs (White) Lumen per package and cost per lumen 2000-2009 6 .922
GaAs Lasers Power density and 1/price 1987-2007 6 .957
Photo-sensors Resolution (1/pixel size) and pixels per price 1987-2007 13 .983
Sensitivity and pixels/price 1987-2007 13 .980
Silicon Solar Cells Efficiency and price per kwH for Silicon 1977-1999 15 .907
Lead Acid Batteries Power per kg and power per cost 1978-1995 12 .99
Power per liter and power per cost 1978-1995 14 .895
108
Capacitors Power per kg and power per cost 1985-2005 7 .992
5.2 Discussion
The strong correlation between improvements in cost and improvements in performance
suggest that the theory behind how performance and cost are improved including the timing of
product and process design changes need to be rethought. One possibility is that product and
process design changes do drive improvements in performance and cost respectively and that
firms are implementing product and process innovations at the same time because they want to
simultaneously reduce costs and improve performance. Thus, it could be that the results are
merely “accidental.” Firms are introducing new products throughout the lifecycle that lead to
improvements in performance while they simultaneously implement process improvements that
lead to cost reductions. However, it seems unlikely that this would be the case for all 17 of the
technologies analyzed or even for the 13 technologies that had correlation coefficients greater
than 0.9.
A second possibility is that increases in some dimensions of performance also lead to
improvements in some dimensions of cost. While users make tradeoffs between cost and
performance at any moment in time87, improvements over time in a technology can change this
tradeoff and thus lead to both improvements in cost and performance88. For example, increases
in the efficiency of solar cells, lighting or the luminosity per Watt of displays or lights, in the
densities of transistors (e.g., Moore’s Law) or memory cells, in the power densities of batteries
or capacities, or in the speeds of electronic devices such as computers can lead to lower costs in
the electricity from solar cells, the cost per lumen of lighting or displays, the cost per power for
capacitors and batteries, the cost per transistor of microprocessors or other ICs, the cost per
memory cell of semiconductor and magnetic memory, and the cost per computation for
109
computers89. Higher efficiencies and greater densities reduce the necessary amount of materials
and the size of equipment and thus costs in addition to increasing performance.
A third possibility is that inter-related product and process design changes are implemented by
either single firms in what may be called “integral” design multiple firms in what many call
modular design or by a single firm in what may be called “integral” design. For example,
although firms such as Intel design both the product and process for new microprocessors,
improvements in many ICs are achieved through cooperation between foundries and design
houses. However, whether they are implemented via modular or integral design, is not the issue
here, it is the types of inter-related product and process design changes that enable the
improvements and in particular the rapid improvements. The next chapter and the analysis of the
next myth, costs fall as cumulative production rises in a learning curve, helps us better
understand the specific types of inter-related product and process design changes that lead to
improvements in both cost and performance.
Appendix
As with the other chapters, we looked for time series data on performance and cost for a
wide variety of technologies and dimensions of performance. This search focused on
technologies experiencing rapid improvements and on those for which measures of performance
exist. We focused on rapid improvements since rapid improvements are the types of
improvements that enable large improvements in productivity and creative destruction
(Schumpeter, 1942)90 and rapid improvements increase the chances that changes in price reflect
changes in cost. Rapid improvements mean that changes in cost will be much larger (orders of
magnitude) than will be changes in profit margins. This caused us to ignore commodities and
other technologies for which performance is not important and for which rapid improvements
are not occurring91. This includes chemicals, agricultural products and materials for which there
is often a fixed chemical composition and thus there is little room for product innovation 92. We
110
agree that for technologies with fixed chemical compositions, improvements in cost are mostly
if not all driven by process innovations.
Technologies with data over many years if not many decades of commercial production are
also desired. This is because a long time frame increases the chances that there will be a change
in emphasis in an industry from performance to cost and thus product to process design.
Whether the change in emphasis comes from the emergence of a dominant design or a change in
the importance of performance and price93, many years of data will increase the chances that
such changes will have occurred, thus causing the purported change from product to process
innovations. Since the main study94 has found that a change from performance to price took 6
years (as did the dominant design) with personal computers, we have looked for technologies
with more years of data points than this and settled on a round number of 10 years.
We looked for such data in a wide variety of sources, as mentioned in previous appendices.
Data on relative performance and relative price were primarily collected since this is the type of
data that is often reported and since relative performance and price are more relevant for this
study than are absolute performance and price. Scientists and engineers measure the
performance of a technology relative to physical variables such as mass, volume, area, chip,
package, or best possible output (i.e., efficiency) since relative performance is more meaningful
than absolute performance. In one case (microprocessor), absolute values for a dimension of
performance (speed) were found and used as “relative performance” in the correlation
calculations.
Relative price rather than absolute price are also reported in the engineering and science
journals. As with relative performance, relative price is more relevant than absolute price for
measuring progress in a technology and it is also more relevant for this paper’s analysis since
analyzing improvements is the goal. For this paper, performance is placed in the numerator so
that the relative measures of performance and cost/price are both moving in the same direction
111
in order that correlation coefficients can be calculated. Thus, while relative performance is a
ratio of performance to a physical attribute such as mass, volume, area, chip, or package, relative
price is a ratio of performance to price.
In summary, we first gathered time-series data on a variety of technologies and on multiple
dimensions of performance if the data could be found. Second, performance-to price ratios were
calculated for the time series if the original data series was not reported in a performance-to
price ratio; this was the case for one dimension of performance for one technology,
microprocessor speeds. Third, correlations were calculated for the two time series, one time
series is for relative performance measures and one is for performance-to price ratios. Since data
is not available for every year in the time series, sometimes averages for adjacent years were
calculated in order to increase the amount of overlap between the two time series and thus the
amount of data in the correlation analyses.
112
Figure 5.2. Improvements in Relative Performance (Left Axis, Blue) and Improvements in Performance-to Price Ratios (Right Axis, Red)
1970 1980 1990 2000 20101.00E+08
1.00E+11
1.00E+14
1.00E+17
1.00E+06
1.00E+08
1.00E+10
a. Speed (Hz) and Transistors/Hz/Dollar for Microprocessors
1970 1980 1990 2000 20100.0001
0.01
1
100
10000
1.00E+08
1.00E+11
1.00E+14
1.00E+17
b. Millions of Transistors and Transistors/Hz/Dollar for Microprocessors
1990 1992 1994 1996 1998 2000 2002 2004 2006 20081
100
10000
1000000
0.000001
0.0001
0.01
1
d. Number of Megabits and Gigabits per Dollar for Flash Memory
1960 1970 1980 1990 20001
100
10000
1000000
100
10000
1000000
100000000
c. Thousands of Bits and Bits per Dollar for DRAM
1940 1960 1980 20000.000001
0.001
1
1000
1000000
1000000000
0.000000001
0.000001
0.001
1
1000
1000000
e. Millions of Computations per kwH and per Dollar for Computers
1950 1970 1990 20100.01
1
100
10000
1000000
0.0001
0.01
1
100
10000
f. MegaBits per cm3 and MegaBits per Dollar for Magnetic Tape
113
1950 1960 1970 1980 1990 2000 20100.0001
0.01
1
100
10000
1000000
0.00000001
0.000001
0.0001
0.01
1
100
g. Megabits per cm2 and GigaBytes per Dollar for Magnetic Disks
1950 1970 1990 20100.0001
0.01
1
100
10000
0.0001
0.01
1
100
10000
h. Data Speeds (gbps) and Speed per M$ and km for Trans Atlantic Cable
1980 1985 1990 1995 2000 2005 2010 20150.01
1
100
10000
1000000
0.000010.00010.0010.010.1110100100010000
i. 1000s of Base Pairs/Person-Day and 1000s of Base Pairs/Dollar for Sequencing Equipment
1915 1920 1925 1930 1935 1940 19450
400
800
1200
1600
2000
0
0.01
0.02
0.03
0.04
j. Watts per Kg and Watts per Dollar for Aircraft Engines
1915 1925 1935 19450
1000
2000
3000
4000
0
0.01
0.02
0.03
0.04
k. Watts per Liter and Watts per Dollar for Aircraft Engines
1890 1910 1930 1950 1970 19901
10
100
1000
0.001
0.01
0.1
1
l. Watts/Kg and Watts/Dollar for Passenger Car Engines
114
1920 1940 1960 1980 200010
100
1000
0.1
1
10
m. Watts/Liter and Watts/Dollar for Passenger Car Engines
1970 1975 1980 1985 1990 1995 2000 2005 20100.0001
0.01
1
100
10000
0.01
0.1
1
10
100
n. Lumens per Package and Lumens per Dollar for Red LEDs
1998 2000 2002 2004 2006 2008 201010
100
1000
10000
1
10
100
1000
o. Lumens/Package and Lumens/Dollar for White LEDs
1985 1990 1995 2000 2005 201010
100
1000
0.0001
0.001
0.01
0.1
1
p. Watts per cm and Watts per Dollar for GaAs Lasers
1985 1990 1995 2000 2005 2010 20150.001
0.01
0.1
1
0.001
0.01
0.1
1
10
q. 1/Pixel Size and Number of Pixels/Dollar for Camera Chips
1985 1990 1995 2000 2005 20101
10
100
0.001
0.01
0.1
1
10
r. Sensitivity (mv/micron2) and Pixels per Dollar for Camera Chips
115
1980 1985 1990 1995 2000 2005 20100.01
0.1
1
10
0.001
0.01
0.1
1
v. Watt Hours per Kg and Watts per Dollar for Capacitors
1976 1978 1980 1982 1984 1986 1988 1990 1992 1994 19960
20
40
60
80
100
0
4
8
12
u. Watts per Liter and Watts per Dollar for Lead Acid Batteries
1978 1980 1982 1984 1986 1988 1990 1992 1994 19960
10
20
30
40
50
0
4
8
12
t. Watt Hours per Kg and per Dollar for Lead-Acid Batteries
1950 1960 1970 1980 1990 2000 20100
10
20
30
0.001
0.01
0.1
1
10
s. Efficiency and KwHours per Dollar for Silicon Solar Cells
116
Chapter 6
The Myth of Costs Falling as Cumulative Production Rises
The fifth myth is that costs fall as cumulative production rises and as improvements are
made to processes on the factory floor. This myth has grown from the large number of
empirical analyses that have been done linking cumulative production and costs in what
many call learning curves. The first such analysis was done by Theodore Paul Wright in 1936
on the costs of fighter jets and many have followed. Kenneth Arrow coined the term, learning
by doing and the early work on learning curves was mostly done on single designs in specific
factories and thus analyzed the impact of factory level changes on factory productivity.
Subsequently, learning curves have been applied to technologies that are manufactured with
new designs and in new factories where the output variable might be cost or performance,
albeit these models are now often called experience curves95.
As noted in the introductory chapter, linking cumulative production to reductions in cost or
improvements in performance in these so-called experience curves can lead to confusion
about how the improvements in cost and performance are being achieved. Are these
improvements being achieved through changes made on the factory floor or through changes
made in laboratories to product and process designs? Many focus on the factory floor due to
the empirical linkage between and cumulative production and falling costs and since it is
easier to focus on factories than on more complicated changes in product and process
designs. Others ignore the question and merely assume that they come from something that is
associated with cumulative production and this by itself is a sufficient answer. Clayton
Christensen’s theory of disruptive innovation96 probably falls into the latter category since
simplicity is often considered a key aspect of a popular theory.
This chapter, along with the analysis in Chapter 5, shows that product and process design
117
changes, typically inter-related ones, are the sources of the improvements. The correlation
that is found between improvements in cost and performance in Chapter 5 suggests that costs
are falling because of changes in product and process design that impact on both
improvements in performance and on cost. Since improvements in processes by themselves
have little impact on performance, eases in cumulative production and changes in process
design by themselves cannot be the reasons for falling costs. This argument is taken one step
further in Chapter 6 by analyzing 13 new technologies that experienced rapid improvements
of greater than 10% per year without commercial production.
6.1 Evidence from 13 new technologies
Table 6.1 summarizes annual rates of improvement for technologies that have
experienced rapid improvements of greater than 10% per year during a time period of zero
commercial production. Like the previous chapters, these technologies are places in various
categories to aid the reader.
Table 6.2 shows data on the time periods (column 3) for which data on rapid rates of
improvement are available, the starts of commercial production (column 2), and recent sales
figures (column 4). These data show there was commercial production during only 17% of
the years (27/156) for which rapid improvements occurred. Table 6.2 (see column 4) also
shows, recent sales figures for each technology and all these figures are smaller than sales of
semiconductors in 1956 ($856 Million in 2013 dollars). Except for printed electronics, which
probably includes technologies other than organic transistors (see below), all of them had
market sizes smaller than did semiconductors in 1954 ($346 Million in 2013 dollars). Thus,
high levels of commercial production still do not exist for any of these technologies.
The small current market sizes are consistent with the dates for the start of commercial
production and together are strong evidence that these technologies were improved with zero
commercial production and still are being improved with small levels of commercial
118
production. While two or three cases might be arguable given uncertainty in the start of
Table 6.1. Technologies with Recent Rapid Rates of Improvement during Periods of Zero
Commercial Production
Technology
Domain
Sub-Technology Dimensions of measure Time
Period
Improvement
Rate Per
Year
Energy
Transfor-
mation
Organic LEDs Luminosity per Watt,
green
1987-2005 29%
Quantum Dot
Displays
External Efficiency, red 1998-2009 36.0%
Solar Cells Efficiency of Organic 2001-2013 12.6%
Efficiency of Quantum
Dot
2010-2013 42.1%
Energy
Trans-
mission
Superconductors:
BSSCO and YBCO
Current-length per cost 2004-2010 115%
Current x length - BSSCO 1987-2008 32.5%
Current x length - YBCO 2002-2011 53.3%
Information
Trans-
formation
Organic Transistors Mobility 1982-2006 109%
Single Walled Carbon
Nanotube Transistors
1/Purity (% metallic) 1999-2011 32.1%
Density 2006-2011 357%
Super-
conducting Josephson
Junction-based
transistors
1/Clock period 1990-2010 20.3%
1/Bit energy 1990-2010 19.8%
Qubit Lifetimes 1999-2012 142%
Bits per Qubit lifetime 2005-2013 137%
Quantum Computers Number of Qubits 2002-2012 107%
Information
Storage
Flash Memory Storage Capacity 2001-2013 46.8%
Resistive RAM Storage Capacity 2006-2013 272%
Ferroelectric RAM Storage Capacity 2001-2009 37.8%
Magneto RAM Storage Capacity 2002-2011 57.8%
119
(Lee, 2005; Sheats et al, 1996; NREL, 2013; Shiohara et al, 2013; Selvamanickam V 2011;
Hasegawa and Takeya, 2009; Franklin, 2013; Fujimaki, 2012; Devoret and Schoeldopf, 2013;
Evans et al, 2011; D-Wave, 2013; ISSCC, 2013)
commercial production, we were conservative (chose earlier if multiple possibilities exist)
about the dates and yet 8 cases are shown in Table 6.1 where production did not start until 9
or more years after the onset of rapid improvement. In summary, these results clearly indicate
that learning by production workers on the factory floor is not a relevant mechanism for these
rapid improvements. The following sections address the mechanisms by which these
improvements occur and they present more detailed data about each of the cases in Table 6.1.
Table 6.2. Starts of Commercial Production and Recent Sales Data
Technology Start of
Commercial
Production
Time Period
for Rapid
Improvements
Recent Sales
Data
($ Millions)
Sources of Sales
Data
Organic LEDs 2001 1987-2005 300 in year 2012 (Display Search
2013)
Organic Transistors 2007 1994-2007 530 (printed
electronics) in
Year 2010)
(Markets and
Markets, 2011)
Organic Solar Cells 2010 2001-2013 4.6 in Year 2012 (IDTE, 2012)
Quantum Dot Solar
Cells
2013 2010-2013 Zero until 2013 (Investor, 2013)
Quantum Dot
Displays
2013 1994-2009 Zero until 2013 (Research &
Markets, 2013)
Resistive RAM 2013 2006-2013 200 (2012) (Yole, 2013)
Ferroelectric RAM 2005 2001-2009
Magneto-resistant
RAM
2004 2002-2011
Phase Change RAM 2006 2004-2012
Single Walled
Carbon Nanotubes
2011 1999-2011 <10 (2011) (BCC, 2012)
120
for Transistors
High Temperature
Superconductor
Wire (YBaCuO and
BiSrCuO)
2006 1987-2008 30 (2011)
30 (2012)
(Connectus,
2012)
Superconducting
Josephson Junction-
based Transistors
2011 1990-2010 First sale of
these
technologies in
2011
(Jones, 2013)
Quantum
Computers
2011 2002-2012
Discussion of the technologies listed in Table 6.1 is covered in five sub-sections (organic
materials, quantum dots, new forms of non-volatile memory integrated circuits, carbon
nanotubes, and superconducting materials). Several of these sub-sections discuss multiple
entries from Table 6.1 since the mechanisms substantially overlap. Each sub-section describes
the relevant dimensions of performance, the impact of these improvements in performance on
cost and gives logarithmic plots of the performance measures against time in figures to
display the time series data. The starts of commercial production are shown on each graph as
a large black arrow and recent sales data is discussed. In addition, each section discusses the
technical changes that drove improvements in performance and cost for specific technologies.
6.1.1 Organic Materials
Organic materials are being used in many electronic applications because they are more
mechanically flexible than are semiconductor materials and because it is potentially cheaper
to fabricate electronic devices with them than with semiconductor materials that require high
temperature processes. These electronic devices include lighting, displays, solar cells, and
transistors. Luminosity per Watt, shown against time in Figure 6.1 is an important dimension
of performance for organic light emitting diodes (OLEDs) in lighting and displays and it also
impacts on the cost for users. Not only does better luminosity per Watt lead to lower cost per
121
lumen97, it also enables smaller devices for a given output and thus lowers material,
equipment, and transport costs. This makes improvements in luminosity per Watt an
important driver and rough surrogate for reductions in the cost of lighting or displays with
OLEDs.
Figure 6.1 shows that about 10 to 100 times improvements were made in luminosity per
Watt before documented commercial production was started in 2001 (see black arrow) for
mobile phone displays. Even after commercial production was started, sales grew slowly with
some ups and downs. For example, the leading market research firm, Display Search, claims
in two different reports that sales were $615 Million in 2008 and $300 Million in 201298.
Thus increases in commercial production do not appear to be a strong explanation for
improvements in cost and performance after 2001 and are even more questionable before this
date where a large fraction of the documented improvements occurred.
1985 1987 1989 1991 1993 1995 1997 1999 2001 2003 20050.1
1
10
100
1000Figure 6.1 Luminosity Per Watt for Organic Light Emitting Diodes
(black arrow = start of commercial production)
Green
Yellow
Blue White
Lum
ens/
Wat
t
Not surprisingly, the literature on OLEDs provides an explanation for the improvements
that different than those based upon improvements on the factory floor. Scientists and
engineers created new types of organic materials that better exploited the phenomena of
electroluminescence. In the literature, one can find graphs similar to Figure 6.1 that identify
122
materials changes at many of the data points. For example, according to such plots in two
sources99, improvements in the lumens per Watt of green, blue, and yellow OLEDs came
from new forms of InGaN, polyfluorenes, phosphorescent materials and molecular solids
such as Tris (8-hydroxyquinolinato) Aluminum.
The second entry in Table 6.1 is another technology that is based on organic materials,
organic transistors. Mobility is a key dimension of performance for transistors since mobility
directly impacts on speed and speed is an important dimension of performance for computers
and other electronic products. Mobility also impacts indirectly on cost since computers are
often evaluated in terms of their cost per instruction100 or cost per processing output101 and
mobility impacts on the speed at which these computers can perform. Thus, the
improvements in mobility shown in Figure 6.2 are often a good surrogate for improvements
in the cost of processing data with organic transistors.
Organic transistors were first commercially used in 2007 to control the pixel values in
flexible electronic paper, which are used in e-books such as the Amazon Kindle102. From
Figure 6.2, improvements of more than 1000 times occurred before the start of commercial
production in 2007. More recent sales data for organic transistors could only be found under
printed electronics, a category that includes any type of electronic circuit that is printed
including transistors and displays. Thus, the $530 Million figure in Table 6.1 is probably a
substantial overestimate for the size of the organic transistor market in 2010 or for more
recent years and increases in production are not a viable explanation for improvements in cost
and performance before 2007 and probably not after that date as well.
123
1980 1985 1990 1995 2000 2005 20100.000001
0.0001
0.01
1
100Figure 6.2 Mobility of Single Crystal and Polycrystalline Organic Transistors (black arrow = start of commercial production)
Singlecrystal
Polycrystalline
Mo
bil
ity
(cm
2 x
sec
)
Like OLEDs, the literature on organic transistors and even the figures displaying the
improvements in mobility in this literature focus on new materials as the sources of
improvements. For example, one paper103 includes the names of organic materials such as
polythiophenes, thiophene oligomers, polymers, hthalocyanines, heteroacenes,
tetrathiafulvalenes, perylene diimides naphthalene diimides, acenes, and C60 alongside the
improvement data. Processes are also important. For example, one scientific paper104 says:
“the search for high mobility materials is still very active. However, the mobility is not only
dictated by the nature of the organic semiconductor; it also strongly depends on other
parameters such as the crystal structure and the quality of the various interfaces that intervene
in the device: interfaces between the insulator and the semiconductor and between the
electrodes and the semiconductor.” Thus, process optimization and materials creation are
inextricably linked for this technology.
The third entry in Table 6.1 is organic solar cells. The efficiency of solar cells is defined
as the percentage of incoming solar energy that is converted into electrical energy and it has a
large impact on the cost of electricity from solar cells. While it contributes equally to
reductions in cost per peak Watt of physical cells as does reductions in cost per area, it
124
contributes more to reductions in the cost per peak Watt than does reductions in the total cost
per area because installation costs depend on the area of the solar cells and better efficiencies
enable smaller areas for a given output105. Like OLEDs, this makes improvements in
efficiency, shown in Figure 6.3, an important driver and useful surrogate for reductions in the
cost of electricity from organic and other solar cells.
1998 2002 2006 2010 2014
Organic
QuantumDots
Figure 6.3 Efficiency of Organic and Quantum Dot Solar Cells (black arrows = starts of commercial production)
25%
5%
0%
Eff
icie
ncy
Organic solar cells were first commercially produced in 2010 by Konarka, a firm that
went bankrupt in mid-2012. One market research firms claims that organic solar cells had
about $4.6 Million in sales in 2011106, the latest year for which we were able to find sales
data. Figure 6.3 shows that before commercial production of organic solar cells began in
2010, significant improvements had been achieved in their efficiency, growing from 3% in
2001 to 8% in 2010 and 11.1% by the end of 2012. (For comparison purposes, single crystal
silicon solar cells have a best laboratory efficiency of 25%). Like OLEDs, the literature on
organic solar cells focuses on the creation of new types of organic materials and creating
these new organic materials often requires new processes. For example, two scientists107
summarize these improvements in the following way: “the development of active layer
125
materials is still the key to boost the efficiency. In order to get better photovoltaic properties,
many properties, like band gap, molecular energy level, mobility, solubility, etc., should be
considered, and how to balance these parameters is the most important part to molecular
design.” Continuing with these two scientists, they describe the creation of new materials
such as blended films of conjugated polymer (electron donor) and small molecular acceptors.
Examples of active materials that enabled improvements in efficiency include polythiophene,
polymers with 2,1,3-Bezothiadiazole pyrrolo derivatives, and bezo-dithiophene-based
polymers.
6.1.2 Quantum Dots
The fourth entry in Table 6.1 is also a solar cell and thus the efficiencies are also shown
against time in Figure 6.3. However, quantum dot solar cells are based on a different physical
principle and a different type of material than are organic solar cells. Quantum dots consist of
small crystals (usually semiconductor crystals) whose size determines their electronic and
optical properties and thus the wavelengths of light that will be absorbed in a quantum dot
solar cell. By varying the size of the dots, it is theoretically possible to create solar cells on a
single layer of material that have efficiencies greater than 80% or more than three times the
maximum theoretical efficiency of conventional solar cells such as those made with organic
or semiconductor materials (including silicon). Like other forms of solar cells, improvements
in efficiency can be considered a surrogate for reductions in costs108.
Commercial production of quantum dot solar cells was reported to have started in late
2013109. The market for all types of quantum dots including displays and medical applications
were only $150 Million in 2011110. Even with significant production, learning from
commercial production is not a good explanation for efficiency improvements such as those
experienced by quantum dot solar cells shown in Figure 6.3. This figure shows that rapid
126
improvements in the efficiency of quantum dots have been recently achieved, rising from 3%
in 2010 to 8.6% in mid-2013. Like organic solar cells, the literature on quantum dot
technologies focuses on the creation of new materials and processes. Semiconductor and
other crystals are grown with new types of materials or by adding new impurities and/or
dopants to these base materials. Types of materials that are mentioned as contributing to the
improvements in efficiencies include conventional semiconductors such as silicon or indium
arsenide, more complex compositions (i.e., alloys), and selenide or sulfides of metals (e.g.,
lead sulfide, lead selenium, cadmium selenium). New structures such as quantum dots that
are grown within dots are also mentioned where new processes are typically required for
these new structures to be effective111.
Quantum dot displays are the fifth entry in Table 6.1. Like quantum dot solar cells, the
size of the dot determines the relevant wavelength of light although in this case, the important
wavelength is of the emitted rather than the absorbed light. Similar to solar cells, this enables
a single layer of material to theoretically emit many different colors and thus result in a lower
cost display than current display technologies. Also like OLEDs, the efficiency with which
electricity is converted to light has an important impact on both the performance and cost of
the quantum dot display. For quantum dots, this efficiency is measured in terms of the
percentage of available electrons that are converted to photons while for OLEDs, the
efficiency is measured in terms of luminosity per Watt. In any case, improvements in the
efficiency can be considered a rough surrogate for reductions in the cost of quantum dot
displays112 and are displayed against time in Figure 6.4.
Commercial production of quantum dots for television displays reportedly began in 2013
by Sony. These quantum dots are used in combination with liquid crystal displays (LCDs) to
increase the range of colors that can be displayed on a television113. As noted above, even the
sales for all types of quantum dots in 2011 was $150 million and these were mostly for
127
biological/medical applications114. Similarly to the technologies previously discussed, the
results indicate that the significant improvements in the efficiency of quantum dot displays
(Figure 6.4 shows 1000 times improvement since 1994) are not due to mechanisms inherent
to production. Instead, the literature on quantum dot displays focuses on the creation of new
materials and typically the same types of new materials, structures and processes that have
enabled improvements in the efficiency of quantum dot solar cells (see above). Additional
mechanisms stress the intricate linking of new materials and new forms of processes such as
layer-by layer assembly methods115 and the use of ZnO nanoparticles and organic layers in
combination with conventional semiconductor-based materials116.
1992 1994 1996 1998 2000 2002 2004 2006 2008 2010 2012
Red
Blue
OrangeYellow
Green
Figure 6.4 Efficiency of Quantum Dot Displays for Different Colors (black arrow = start of commercial production)
10%
1%
.1%
.01%
Eff
icie
ncy
6.1.3 New Forms of Non-Volatile Memory Integrated Circuits
The next four entries in Table 6.1 are for new forms of memory ICs and their
improvements are driven by a different set of technical changes than are the previously
discussed technologies. These four entries are for different types of RAM (random access
memory) ICs that can be defined as non-volatile memory (NVM). NVM refers to memory
that retains its value when the power is switched off. The most familiar type of NVM is
128
called flash memory and it is familiar to many of us because it is used in mobile phones. It
enables our phones to remember our phone numbers, music, and videos even when the power
has been switched off.
Like all forms of chip-based memory, a key dimension of performance for NVM is
storage capacity. The performance of semiconductor memory is usually measured in terms of
the number of bits per chip and these increases are typically achieved by increasing the
number of bits per area. Increases in the number of bits per area usually lead to lower costs
per bit because the higher densities lead to lower material and equipment costs per bit, along
with enabling faster speeds. This is also the case with Moore’s Law, which is typically
discussed for microprocessors, but the analysis holds as well for other types of integrated
circuits such as RAM and other types of memory ICs. Increases in the numbers of transistors
or memory cells per IC chip lead to lower cost and higher speeds117 and similar improvements
will occur with NVM. The four entries on NVM refer to four new types of NVM that are
being developed as potential replacements for flash memory. Although these four types of
NVM are based on different physical principles, materials, and structures and different ones
from flash memory (which uses silicon), patterns on them are fabricated using some of the
same processes and equipment that are used to fabricate patterns on ICs including flash
memory ICs.
As shown in Figure 6.5, substantial improvements have been achieved in these four types
of NVM some of which were achieved before commercial production for each alternative
(shown by the 4 black arrows) was started. This is particularly true with resistive RAM, for
which an improvement of 100 times occurred before commercial production started in 2013.
Since the total commercial sales for all four forms of NVM were only $200 million in
2012118, and since most improvements occurred before the start of commercial production, it
appears again that mechanisms associated with commercial production-based learning have
129
not been a major mechanism for the improvements even after commercial production was
started. Instead, most of the improvements were achieved in laboratories where prototypes
were fabricated119.
The main mechanism for achieving the improvements in storage capacity has been from
reducing the feature sizes associated with the memory cells. Differences in storage capacity
between different forms of NVM are largely a function of differences in these feature sizes 120.
This is the same mechanism by which the number of transistors per chip is increased in
conventional ICs such as flash memory, DRAM, and microprocessor ICs. Firms such as Intel
reduce the size of the features that define a transistor or memory cell and thus are able to
increase the number of transistors or memory cells per chip121. In the case of the new forms of
NVM, the reductions in feature size are made largely by the appropriate modification of
equipment and processes that are borrowed from the manufacture of flash memory or other
integrated circuit-related industries.
2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013
0.0001
0.01
1
100
Phase Change RAM
Ferro Electric RAM
MagneticRAM
Resitive RAM
Figure 6.5 Number of Memory Bits (Gb) per RAM (Random Access Memory) Chip (black arrows = start of commercial production)
Sto
rage
Cap
city
per
Chi
p (G
b)
6.1.4 Carbon Nanotubes for Transistors
130
Carbon nanotubes (CNTs) are another technology that is being developed for ICs largely
because of their very high electrical and thermal conductivities. CNTs are composed solely of
carbon atoms, just as graphite, diamond, and graphene are. While graphene is a one-atom
thick layer of carbon atoms, one can think of CNTs as graphene that is rolled into cylindrical
tubes with either open or closed ends. These CNTs can be produced with single, “few,” or
“multi” walls of which the single wall ones have the highest performance and cost. The
market for single walled CNT has reached $10 Million and the market for all three types of
carbon nanotubes had reached $180 Million by 2011122.
1995 1997 1999 2001 2003 2005 2007 2009 2011 2013 20150.01
0.1
1
10
100
0.1
1
10
100
Density with InconsistentFeature Size
Density withConsistentFeature Size D
ensi
ty (
Car
bon
Nan
otub
espe
r m
icro
met
er)
Pur
ity (
% C
onta
min
ant)
Figure 6.6 Purity (left axis) and Density (right axis) of Carbon Nano Tubes for Tran-sistors. Density is for Consistent and Inconsistent Feature Size
Purity
Start of Production
For transistor applications, single-walled CNTs are needed that have both high purity and
density123. Figure 6.6 shows that both of these dimensions of performance have been
improved at rapid rates and most of these improvements were achieved before commercial
production of single-walled CNTs by firms was begun in 2011 (see black arrow).
Improvements in the purity of the CNTs have been achieved by improvements in processes.
As described by one paper124, this includes “post-synthetic efforts to purify and sort carbon
nanotubes by their physical and electronic structure” and the “selective growth of carbon
131
nanotubes with predetermined properties.” In another paper125 the writes, “Jin and colleagues
have managed to achieve the selective removal of metallic CNTs from an array of such
nanotubes on a chip without damaging the semiconducting nanotubes.”
Increases in the density of CNTs have come from improvements in processes and new
materials. This includes the growth of CNTs on new types of substrates such as quartz and by
coating the CNTs with certain molecules to tune their attraction to different surfaces126.
6.1.5 Superconductors
The final three technologies in Table 6.1 are superconductor-related. As their name suggests,
superconductors conduct electricity with zero resistance and this enables them to be very
effective conductors of electricity and to enable the creation of strong magnetic fields. The
performance of superconducting materials is typically measured in terms of the highest
temperature at which superconducting occurs (the critical temperature), the amount of current
and magnetic field they can support before superconducting disappears, and these dimensions
in combination with length due to the difficulties of fabricating long superconducting wires.
Higher temperatures, currents, and magnetic fields are also related to cost since higher
temperature superconductors reduce cooling costs and higher currents and magnetic fields
reduce the amount of necessary materials and thus their costs127.
More than 33 superconducting materials have been created128 of which the highest recorded
critical temperature is 153 degrees Kelvin. Five of these materials are capable of
superconducting at temperatures higher than 77 degrees Kelvin and are often called “high-
temperature” superconductors. These higher temperatures enable the replacement of liquid
hydrogen with liquid nitrogen as a method of cooling thus reducing the cost of cooling.
Furthermore, many of these newly created materials can handle higher currents or magnetic
fields at a specific temperature than can previously created materials.
132
Although the overall production of superconductors has grown steadily over the last ten
years reaching $4.5 billion in 2007, few of these sales are for superconducting wires or rapid
single flux Josephson Junctions of which the latter are used in quantum computers. Instead,
the largest application for superconductors is magnetic resonance imaging in which so-called
“low temperature” superconductors are fabricated into magnets. Electric power applications
such as cables, transformers, motors, and generators that use high temperature
superconductor wires only represent a few percent of the superconductor market129. These
wires are fabricated from high temperature superconductors such as the ones (BiSrCaCuO
and YBaCuO) that have experienced rapid improvements in the cost per kilo-amp-meter
(Figure 6.7) or in the current times length (Figure 6.8). The market for high temperature
superconductors was only $30 Million in 2011 with commercial production beginning in
2006130. Thus, mechanisms relating to production experience are probably not important
sources of these significant improvements similar to the previously discussed technologies in
this paper.
2003 2005 2007 2009 2011100
1000
10000
100000
Figure 6.7. Cost per kiloamps-meter for Superconducting Cable(black arrow = start of commercial production)
$/ki
lo m
ps-m
eter
133
1985 1990 1995 2000 2005 2010 20151
10
100
1000
YBaCuO
BiSr CuO
Figure 6.8. Current (Amps) x Length (km) for Two Types of Superconducting Cables ((black arrow = start of commercial production)
Am
ps x
Len
gth
According to the engineering literature, improvements in the current carrying capability and
cost of BiSrCaCuO and YBaCuO (Figures 6.7 and 6.8) were achieved largely through
improvements in processes but also through modifications to the materials that are used to
package these materials into wires. For increasing current times length, achieving an in-plane
grain alignment of the material’s crystals was important and this was achieved with a new
process (ion beam assisted deposition) and a new substrate, a rolling-assisted biaxially
textured one. For reducing the cost of YBaCuO, one challenge was to reduce the content of
silver due to its high price while retaining the wire’s strength. This was achieved by
combining YBCO with nickel and other dopants131.
Rapid single flux quantum (RSFQ) Josephson junctions are another application for
superconductors in which the rapid improvements shown in Figures 6.9, 6.10, and 6.11 have
been achieved prior to commercial production. Named for their discoverer, Brian David
Josephson, Josephson junctions consist of a thin non-superconducting material that is
134
sandwiched between two superconducting materials and for which quantum tunneling can
occur across the non-superconducting material. These junctions can be used to construct
various electronic devices such as single-electron transistors, qubits, superinductors,
superconducting quantum interference devices, superconducting tunnel junction detectors,
and rapid single flux quantum (RSFQ); RSFQ is the device of interest in this section. RSFQs
that are constructed from these junctions are orders of magnitude faster and use orders of
magnitude less power than do conventional ICs. Since the cost of computing is often
measured in cost per instruction and cost per energy132, the improvements in the speed of
RSFQ also impacts on the cost of computing, as does improvements in power consumption.
Thus, improvements in speed and power consumption can be considered surrogates for
reductions in cost.
Improvements in speed, i.e., clock period, and power consumption, i.e., bit energy, (See
Figure 6.9) were being achieved at a rapid rate before commercial production began in 2011
(see black arrow) for quantum computers133. Like the NVM and conventional ICs that are
discussed above, these improvements were achieved by reducing the feature size of the RSFQ
Josephson junctions. Reducing the size of the RSFQ Josephson junctions reduced the
distance to be traveled by electrons and these reductions in feature size enabled both
increases in speed and reductions in power consumption134.
135
1990 1995 2000 2005 2010 201510
100
1000
0.001
0.01
0.1
1
Figure 6.9 Bit Energy (left axis) and Clock Period (right axis) for Super- conducting Josephson Junctions (black arrow = start of commercial production)
Bit
Ene
rgy
(Fem
to J
oule
s)
Clo
ck P
erio
d (P
ico
Sec
onds
)
Clock Period
Bit Energy
1998 2002 2006 2010 20140.001
0.1
10
1000
100000
Relaxation Time
Coherence Time
Cavity Lifetime
Figure 6.10 QuBit Lifetime for Several Definitions of "Lifetime"(black arrow = start of commercial production)
Lif
etim
e (n
anos
econ
ds)
One application for these RSFQ Josephson Junctions is quantum computers. Quantum
computers differ from conventional computers in that bits can be in “superposition,”
representing 0 and 1 at the same time according to a probability distribution. The bits in a
quantum computer are called qubits and by coupling multiple qubits, the performance of a
quantum computer rises at a much faster rate than do increases in the number of qubits. While
conventional computers operate on a base two system, i.e., 0 or 1, and thus performance rises
136
linearly with increases in the number of bits, the performance of quantum computers rises
non-linearly as the number of qubits are increased135. The problem for quantum computers is
that “keeping qubits in superposition long enough to do anything useful with them has proven
very hard”136.
Nevertheless, this problem is gradually being solved as improvements in Qubit lifetimes and
in the number of bits per lifetime have been achieved as shown in Figures 6.10 and 6.11. The
number of bits per lifetime (Figure 6.12) is equivalent to the number of measurements, each
with one bit of precision that would be possible before an error occurs. These improvements
have been achieved by creating new types of qubit structures and new processes for making
these structures. The new structures include different sizes and orientations of tunnel
junctions, superinductors, and resonators and new processes include exposure to microwave
radiation. These approaches are given unusual names such as Quantronium, Fluxonium,
Transmon, and improved Transmon137.
2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 20141
10
100
1000
10000
Figure 6.11 The Number of Bits per QuBit Lifetime(black arrow = start of commercial production)
Num
ber
of B
its
The improvements in Qubit lifetimes and in the number of bits per lifetime have enabled
the number of Qubits in a quantum computer to be recently increased more than 100 times as
137
shown in Figure 6.12. The figure demonstrates that much of these improvements were
achieved without the commercial production of quantum computers. The first prototype was
constructed in 2002 and commercial production of both quantum computers and RSFQ
Josephson junctions started in 2011 with the first sale of a quantum computer138.
6.2 Interpretation of Results
The 13 technologies listed in Table 6.1 were shown in the previous section to have achieved
rapid improvements in performance and/or cost during periods of zero commercial
production and all of them still have low levels of commercial production. Some of the cost
reductions are known because cost data was collected while other cost reductions are inferred
based on the fact that improvements in some dimensions of performance are equivalent to
reductions in cost for a given performance.
2000 2004 2008 20121
10
100
1000
Num
ber
of B
its
Figure 6.12 Number of Qubits in Quantum Computers (mostly prototypes)(black arrow = start of commercial production)
The findings detailed in the previous section demonstrate that the rapid reductions in cost
and increases in performance are - as expected since the cases were chosen to avoid
production learning - not due to mechanisms associated with factory floor activities and
138
production experience. One predominant mechanism found in these cases was creating
materials that better exploit a physical phenomenon where these new materials often required
new processes. New materials were important for OLEDs, organic transistors and solar cells,
quantum dot solar cells and displays, and superconductors. Scientists and engineers created
organic materials that better exploited the phenomenon of electroluminescence for OLEDs,
the photovoltaic phenomenon for solar cells, and the semiconducting phenomenon for
transistors. They created semiconductor and other materials that better exploited the
phenomenon of quantum dots and other materials that better exploited superconductivity.
Sometimes, the focus was on a single layer of active material while other times it was for a
combination of different materials where each layer in the combination may have been
tweaked with impurities and dopants. The multiple ways in which new materials are created
for a single technology is consistent with the emergence of bottlenecks in a system of
materials or components139.
A second mechanism for improving performance and cost in the thirteen cases was
improvements in processes, which as noted in the preceding paragraph, is a subset of the first
method since creating new materials often required new processes for the creation or
improvement of the material. Nevertheless, improvements in the performance of a single
material often involved new processes where these new processes involved slight changes to
the material composition, including the addition of impurities or dopants. Our finding of the
importance of process research is consistent with other research140 that identified new
processes arising from research as a major source of cost reductions. For carbon nanotubes,
the efforts were aimed at improving their purity and density. This includes post-synthetic
efforts to purify and sort carbon nanotubes by their physical and electronic structure, the
“selective growth of carbon nanotubes with predetermined properties, and the growth of
CNTs on new types of substrates such as quartz ones and by coating the CNTs with certain
139
molecules to tune their attraction to different surfaces141.
A third mechanism (in addition to materials creation and new processes) for improving
performance and cost is reducing the scale of features, a mechanism that has perhaps received
too little attention in the economics and innovation literature, albeit there are exceptions142.
The scale of features was reduced to increase the storage density in new forms of non-volatile
memory (RAM) and to increase the speeds of superconducting rapid single flux quantum
(RSFQ) Josephson Junctions. This involved reducing the size of the memory cells and RSFQ
Josephson Junctions, i.e., changes to the product design, and introducing processes that
enabled the better control that is necessary to achieve these smaller feature sizes in the
product design. Improvements in the performance of RSFQ Josephson Junctions in the form
of faster speeds and lower power consumption also contributed to improvements in the
performance and cost of quantum computers.
6.3 Discussion
The predominant viewpoint is that costs decrease as cumulative production rises 143 in a so-
called learning curves. The early work on learning curves focused on single designs that were
made in single factories but later work has considered technologies that involve multiple
designs and multiple factories are involved. As noted several times in this book, linking
cumulative production to reductions in cost or improvements in performance in these so-
called experience curves can lead to confusion about how the improvements in cost and
performance are being achieved. Many focus on the factory floor due to the empirical linkage
between and cumulative production and falling costs and since it is easier to focus on
factories than on more complicated changes in product and process designs.
This chapter has shown that some technologies experience rapid improvements in cost and
performance with no commercial production. This clearly suggests that costs can be reduced
without increases in cumulative production and that factory floor activities are not the sources
140
of the cost reduction. It showed that most of the improvements came from changes in product
and process designs that are being developed in laboratories.
These product and process design changes are consistent with the analysis done in Chapter
5. Creating new materials to exploit physical phenomena involve inter-linked product and
process design changes. The creation of new materials often requires new processes and these
new materials will often lead to improvements in both performance and cost. Improvements
in luminosity per Watt of new lighting technologies, of the efficiency of solar cells, and the
current density of superconductors also led to reductions in cost since fewer materials and
thus lower equipment and transport costs were required.
Reductions in the scale of features also required inter-linked product and process design
changes. Smaller memory cells, superconducting Josephson junctions, and transistors enable
faster speeds since the electrons have shorter distances to travel. They enable lower costs
since more memory cells, junctions, and transistors can be placed in a smaller area and thus
can be processed with the same size equipment and require fewer materials.
As for the technologies that primarily experienced improvements in processes, these are not
the improvements that learning curves emphasize and they are consistent with results of
Chapter 5. These process improvements are done in laboratories and they involve
technologies in which the formulas are mostly fixed; these technologies were ignored in
Chapter 5. Nevertheless, some fixed formula technologies do have performance measures of
which purity and quality of crystals are common ones and improvements in processes are
necessary for these improvements in performance.
One important question that was not addressed in this chapter was the extent to which
sources of cost reductions change as cumulative production rises over time. We would expect
that as commercial production increases, the importance of larger scale equipment becomes
greater. But how much greater? The analysis in Chapter 4 suggests that increases in the scale
141
of production equipment do not lead to the rapid rates of improvement that exist for the
technologies analyzed in this chapter. In this chapter, for the six of the technologies for which
we have data, we observe no great changes (either accelerations or decelerations) in
improvement rates (see Figures 6.1 through 6.12) after commercial production begins, which
is at least suggestive that combinations of materials creation, process creation and reductions
in scale continue to impact on cost and performance in these domains as the levels of
commercial production increase. We address this and related issues in the next chapter.
A final point to make is that a lack of accelerations or de-accelerations is also consistent
with Chapter 2’s finding that performance vs. time curves more closely resemble straight
lines on a logarithmic plot than the classical S-curve. Chapter 2 only found one technology
that experienced the first half of the S-curve. The results from both Chapter 2 and this chapter
suggests that early improvements are a signal for a technology’s potential for improvements
and thus can be used to better understand the potential for a new technology. Technology’s
with rapid rates of improvements will likely continue to experience these improvements
while technologies with slower rates of improvement may not ever experience rapid
improvements.
Appendix
We looked for new technologies that have experienced rapid improvements in cost or
performance (>10% per year) during time periods in which there was no commercial
production. As a point of comparison we note that integrated circuits (ICs) have experienced
improvements in the number of transistors per chip of greater than 30% per year - commonly
known as Moore’s Law. We looked for technologies that are experiencing rapid
improvements because technologies with rapid improvements are more likely to have a large
impact on improvements in productivity and lead to the emergence of technological
discontinuities and creative destruction than are technologies with less rapid improvements.
142
We also note that annual rates of improvement are very different from rates of improvement
for each doubling in cumulative production, which many measure144 They are only equivalent
when cumulative production is doubling each year. Since most technologies experience unit
or sales growth rates of much less than 100% a year, annual rates of improvement are often
1/3 to 1/10 those for each doubling.
We first looked for technologies that are currently experiencing rapid improvements for
two reasons. First, it is less likely that improvements occur with no commercial production
for older than for newer technologies. Increases in university R&D over the last 50 to 100
years (Murray, 2002, 2004) have increased the chances that improvements will occur for
example before the beginning of commercial production. When there was little university
R&D before for example WWII, pre-commercial production improvements depended on the
research activities of corporations. Since corporations have much shorter time horizons than
do university and government laboratories, it is likely that there is more pre-commercial
research done now than 50 years ago. Second, publication in journals is much less likely for
corporate work than university work so it is easier to find data for newer than older
technologies. Thus, the paper focuses on newer technologies- specifically technologies that
have experienced rapid improvements in the last 10 years.
We looked for time series data in the leading scientific journals such as Nature, Science,
Optics, and Nanoletters, annual reports by reputable scientific organizations such as the
International Solid State Circuits Conference, general technology websites, technology-
specific web sites for new technologies such as superconductivity and quantum dots, and in
searches on Google. For the latter, we searched for data on technologies that are mentioned in
“science-based” journals with terms such as cost, performance, trends, improvements, and
specific dimensions of performance.
The search and data analysis led to the creation of a relatively large data set on
143
technologies and their rates of improvement, which are summarized in Table 6.1. All of these
technologies are based on a new concept or paradigm145 and thus Table 6.1 does not include
new and better versions of memory ICs, computers, liquid crystal displays, or mobile phones,
all of which can be defined as new product (and process) designs but probably not new
technologies that have been “invented.” The technologies are placed into 7 categories, the
first six of which are the transforming, storing, and transporting of energy and information,
which is consistent with characterizations of engineering systems146. In total, there are 30
individual technologies shown in Table 6.1. Since a variety of performance measures are
often relevant for a specific technology, data was collected on multiple dimensions some of
which are represented in performance of basic functions per unit cost while others are in
performance of functions per mass or per volume totaling 45 metrics in the table.
Second, within the technologies listed in Table 6.1, we looked for technologies that
experienced improvements with no commercial production. To examine this question, we
gathered data on the start of commercial production and data on the current level of
sales/production. This data was searched for on the Internet. Web searches were done for
specific names of technologies along with keywords such as commercial production, start of
production, and sales. Commercial production is defined as the production of a technology by
firms for a specific application and this does not include production by universities or the
development and construction of prototypes by firms. In gathering this data, we looked for
data in consulting reports published by organizations such as Dataquest, BCC Research, Lux
Research, ICE (Integrated Circuit Engineering), and DisplaySearch and we were careful to
emphasize actual rather than forecasted sales. Consulting reports often focus on forecasts
rather than actual sales because most of their readers are more interested in the future than in
the past.
We then selected technologies for which our performance data showed rapid
144
improvements before the start of commercial production according to the production/sales
data. As a check on the start of commercial production date, we gathered recent sales data on
each technology. Technologies with recent start dates for commercial production would not
be expected to be currently experiencing large sales. For comparison purposes, we note that
the market for semiconductors was $40 million in 1954 and $100 million in 1956 (Tilton,
1971) or $346 million and $856 million respectively in 2013 dollars. The years 1954 and
1956 were three and five years respectively after the first transistors were introduced in 1951.
Third, for the selected technologies, we looked for information on how the improvements
occurred, i.e., the mechanisms. Since each technology experiencing rapid improvements
during a time period of no commercial production is based on a single concept or paradigm,
changes in the concept or paradigm are not the sources of the improvements. Thus, other
types of technical changes were the sources of these improvements. We investigated the
following questions. Were these changes made to the product and/or process design; what
types of product and process design changes occurred? Some argue that changes made to the
physical scale of a product design impact on cost and performance147. But how did these
changes in scale impact on cost? Were new materials and new material classes involved?
Were these product changes completely independent of process design changes148 or were
they inter-related149? Did the bottlenecks change as the improvements proceeded150?
Fourth, we looked for information on the relationship between performance and cost and in
particular, how improvements in some measures of performance can lead to reductions in
cost. While users make tradeoffs between cost and performance at any moment in time,
improvements over time in a technology can change this tradeoff and thus lead to both
improvements in cost and performance151. This interaction of improvements in cost and
performance can lead to reductions in cost essentially arising from improvements in
performance. For example, increases in the efficiency of solar cells (US DoE, 2010), lighting
145
(Azevedo et al, 2009), or displays (e.g., luminosity per Watt), in the densities of transistors or
memory cells (e.g., Moore’s Law), or in the speeds of electronic devices (Koh and Magee,
2006) can lead to lower costs in the electricity from solar cells, the cost per lumen of lighting
or displays, the cost per transistor of microprocessors, and the cost per computation for
computers152.
To identify the relationship between cost and performance and how these improvements
occurred, we analyzed many technical/scientific articles. We looked for discussions of the
improvements and the other questions mentioned in the previous paragraph. The sources of
the time series data were often good sources for information on the mechanisms
improvements. This was particularly true with “science-based” journal papers. In all cases,
however, we searched multiple papers, articles, and technical reports in order to obtain a
deeper understanding of the improvements and to test any hypotheses that were generated by
reading the “Science-based” Journal articles. The mechanisms were only identified by this
inductive process and were not pre-selected.
Chapter 7
The Reality of Technology Change
Chapter 7 presents a reality of technology change that is based on the analyses from the
previous chapters. Different technologies have different rates of improvement primarily
because of differences in opportunities for improvements from changes in product and
process designs. Some technologies experience more rapid rates of improvement because
they benefit more from the two inter-related product and process design changes described in
146
the last chapter than do other technologies. These rapidly improving technologies have a
greater chance of becoming economically feasible than do other technologies for either the
first application or an increasing number of applications and they have a higher chance of
impacting on higher-level systems.
Previous chapters also found that performance vs. time curves more closely resemble
straight lines on a logarithmic plot than an S-curve. Accelerations do not occur and physical
limits take decades if not longer to emerge. The improvements in the new technology are not
driven by a slowdown in the old technology and in fact multiple technologies are competing
to replace a new technology even without a slowdown in the old technology. This
competition primarily occurs in laboratories where advances in science and accidental
discoveries determine the timing of the first recorded performance metrics and the rates of
improvement largely determine the eventual winning technologies. A lack of S-curves means
that pre-production early rates of improvement can provide an important signal for future
rates and thus help us identify the technologies with the large potential for rapid
improvements.
This reality of technology change is described in two ways. The first way is to discuss
technology change from the standpoint of invention and commercialization. How do new
technologies proceed from invention to commercialization and what does this mean for firms,
startups, universities, and public policy? The second way is to show how the two types of
product and process design changes impact on improvements in cost and performance and
cost for a larger number of technologies than were described in Chapter 6. This includes
higher-level systems, the impact of component product and process design changes on these
higher-level systems and the search for new higher-level systems that are made possible by
improvements in components.
147
7.1 Invention and Commercialization
Many discussions of technology change distinguish between different phases in a so-called
lifecycle for technologies. Some distinguish between basic research, applied research and
development where new technologies may proceed through these phases or at least build
from previous ones. Others distinguish between science, technology and commercialization
where new technologies are a physical product, service, or artifact that often build from
advances in science. These advances in science often involve new physical phenomena or
better and more complete explanations for existing ones153. Still others distinguish between
the invention, commercialization and diffusion of new technologies.
The early phases of this lifecycle are the most interesting phases and also the ones that are
least understood. Existing descriptions focus on how inventions are introduced and refined
both before and after they are commercialized. This is a recursive process in which
combinatorial search154 is done and in which advances in science may support this
combinatorial search. Recursion occurs in the development of concepts and their translation
into both working prototypes and economically feasible products. For working prototypes,
problems and sub-problems are recursively solved often at the system and component levels
until a working prototype emerges155.
This book’s focus on performance and cost enable us to characterize this process of
recursion and combinatorial search in terms of performance and cost and thus provide more
details than do the above-mentioned descriptions of phases. Improvements in performance
and cost are essential for a new technology to become economically feasible and performance
and cost data enable us to compare new technologies with old ones in various markets and
thus better understand how and when they become economically feasible. This book’s
research suggests that the improvements in cost (and performance) are achieved in a
continual manner in which plots of costs and performance more closely resemble a straight
148
line on a logarithmic plot than an S-curve. Increase in demand or slowdowns in the old
technology do not cause an acceleration in the rates of improvement and instead the rates of
improvement are relatively constant over many decades. Furthermore, multiple new
technologies compete with each other and with the old technologies and performance and
cost data helps us understand which new ones have a greater chance of becoming
economically feasible and when this might occur.
This book’s research also suggests that time or research effort might be a superior
independent variable to cumulative production since it more effectively explains pre-
production improvements. Chapter 6 suggested that many improvements occur before
commercial production begins and thus time is more relevant than cumulative production
during the early phases of a new technology. Furthermore, many higher-level systems also
experience improvements in performance and cost before cumulative production begins since
their improvements are primarily occur through improvements in components. This can cause
new forms of high-level systems to suddenly appear as if from nowhere when in reality they
were gradually becoming economically feasible as their components experienced relatively
constant rates of improvement. Here the challenge is to envision the types of new systems
that might benefit from improvements in components and that might appeal to users.
Recursion and recombinant search156 are an important part of this continuous process of
improvements and this search includes the revision of concepts during invention and beyond.
Firms search for better product and process designs both during and after commercial
production begins. Concepts, prototypes and process designs are redesigned over man years
in a recursive process.
Two types of recursion in product and process designs appear to be very important:
creating new materials and changes in scale. Creating new materials includes new processes
and when the materials are fixed, changes in processes are a subset of this method of
149
recursion. Changes in scale include both increases and reductions in scale. Both these
methods of recursion add more detail to how technologies experience improvements in cost
and performance and thus proceed from invention to commercialization and diffusion. What
are the relative importance of these methods of recursion, does their importance depend on
the type of technology, and does this importance, change over the life cycle of a specific
technology? The simple answers to the last two questions are yes.
Different technologies experience improvements in different ways and at different rates and
the mechanisms by which technologies experience improvements change over time. Based on
the analyses in Chapters 4 and 5, factory level learning and increases in the scale of
production equipment cannot explain most of the improvements in cost in performance.
Although increases in the scale of production equipment are more important for some
technologies, e.g., chemicals and materials157 than are others, chapter 4 showed that chemicals
have a much lower rate of improvement than do other technologies although they benefit
from increases in scale more than do other technologies. The correlations that Chapter 5
showed between improvements in cost and improvements in performance also suggest that
process design changes by themselves, either increases in scale or learning in factories,
cannot be important reasons for rapid improvements in cost and performance. Some
combination of product and process design changes are occurring that enable improvements
in both performance and cost and this is even truer for rapid rates of improvement.
The relative importance of these product and process design changes depends on the
technology and their rates of improvements. For chemicals and other fixed formula
technologies, improvements in processes and increases in the scale of production equipment
are more important than are the above-mentioned changes in product and process design. For
other technologies, the relative importance of these mechanisms depends on the rates of
improvement. Technologies with slow rates of improvement may benefit more from simple
150
learning in factories and increases in scale of production equipment. But for technologies
with rapid rates of improvement, increases in scale of production equipment and learning in
factories cannot explain the improvements.
For these rapidly improving technologies, the mechanisms identified in Chapter 6
continue to be important after commercial production begins and increases for many
technologies. This is largely because increases in cumulative production and demand also
induce increases in R&D spending158 and this R&D spending enables new product and
process designs to be developed and implemented. The key issue is the ease of achieving
improvements. This explains the early and later rates of improvement; and thus the extent to
which firms can benefit from their increases in R&D spending. If the improvements are easy
to achieve, then R&D spending will increase and the improvements will continue at a
constant rate.
This logic may also apply to the pre-commercial improvements where increases in R&D
spending are made by firms, universities and government funding agencies in responses to
their perceptions about the ease of achieving improvements. If they perceive that there is a
high potential for achieving improvements, they are more likely to continue to increase R&D
spending. Furthermore, this logic may also apply to researchers who will invest more of their
time and effort in technologies that they perceive to have high potentials for achieving
improvements.
As commercial production begins and increases, factory floor learning and increases in the
scale of production equipment will become more important. However, for technologies that
experience rapid rates of improvement, learning in factories can only explain a small fraction
of the improvements. Although learning in factories is highly emphasized by the economics
and management literature (references), both Chapter 4 and 5 suggest that this learning is less
important than ordinarily thought. It is probably primarily important when new product and
151
process designs are implemented. Each time these new designs are implemented, various
production problems must be solved perhaps multiple times as the scale of production is
increased from pilot to full production (Mathews and Cho, 1999; Hatch and Mowery 1998)
and this problem solving closely resembles the learning by doing on factory floors that is well
covered in the literature. But this learning merely enables firms to achieve the full benefits of
the new product and process design changes.
This chapter now describes the changes in product and process designs in more detail in
order to better understand how they apply to some technologies more than other technologies.
It covers creating new materials and reductions and increases in physical scale, beginning
with creating new materials. It does this by describing how these types of recursion impact on
a longer list of technologies than was examined in Chapter 6; data for many of these
technologies were presented in previous chapters.
7. 2 Creating Materials that Better Exploit Physical Phenomena
The technologies shown in Table 7.1 primarily experience improvements through the
creation of materials that better exploit physical phenomena where there is a tight linkage
between creating materials and the processes for making them. Some of these materials can
be defined as new classes of materials (See Table 7.2) while others can be defined as
modifications to existing materials. Beginning with lighting and LEDs (light emitting diodes)
at the top of Table 7.1, scientists and engineers improved the luminosity per Watt by finding
materials that better exploit the phenomena of incandescence, fluorescence, and
electroluminescence. For example, they found new combinations of semiconducting
materials such as gallium, arsenide, phosphorus, indium, and selenium for LEDs and new
combination of organic materials such as nitrides and polymers for organic LEDs (OLEDs).
Many of the improvements in semiconductor LEDs also led to improvements in
152
semiconductor lasers, due to the similarities between them while many of the improvements
in organic LEDs are leading to improvements in OLED-based displays. Furthermore, better
lasers also required better materials for the heat sinks, mirrors, and bonders that are part of a
laser package159.
Finding or creating new materials also lie behind much of the improvement in the
performance and cost of photo-sensors (i.e., camera chips), solar cells, load bearing and
magnetic materials, energy storage devices, and crop yields. Scientists and engineers created
new types of materials and processes for them that more effectively translate photons into
electrons for photo-sensors and solar cells, which exploit the photoelectric and photovoltaic
effects respectively and depend on improvements in processes. For photo-sensors, new
materials and processes were created for the lenses, electrodes, and filters in them. For solar
cells, in addition the new classes of materials that are shown in Table 2, New materials and
processes for them were also created for load bearing materials such as composites, and
magnetic materials such as rare earth ones. For energy storage devices, new materials were
created for the anode and cathode of batteries that better convert chemical into electrical
energy; these include various forms of lead, nickel-metal hydride and lithium. New materials
were also found for energy storage with capacitors and flywheels. For crop yield, new
biological materials such as seeds in addition to chemical ones such as fertilizers, pesticides,
and herbicides were created160.
7.3 Geometrical Scaling: increases in scale
The concept of geometrical scaling can help us understand the technologies that will benefit
from increases or reductions in physical scale. Geometric scaling refers to the relationship
between the geometry of a technology, the scale of it, and the physical laws that govern it. Or
as others describe it: the “scale effects are permanently embedded in the geometry and the
153
physical nature of the world in which we live161.”
Table 7.1 Creating Materials that better Exploit Physical Phenomena
Technology
Domain
Sub-
Technology
Dimensions of measure Time
Period
Improvement
Rate Per
Year
Energy
Trans-
Formation
Lighting Light intensity per unit cost 1840-1985 4.5%
Light emitting
diodes (LEDs)
Luminosity per Watt 1965-2008 31%
Organic LEDs Luminosity per Watt 1987-2005 29%
GaAs Lasers Power/length-bar 1987-2007 30%
Photosensors Light sensitivity 1986-2008 18%
Solar Cells Power output per unit cost 1957-2003 16%
Energy storage Batteries Energy stored per unit mass 1882-2005 4%
Capacitors Energy stored per unit mass 1962-2004 17%
Flywheels Energy stored per unit mass 1975-2003 10%
Living
Organisms
Biological
transfor-
mation
US wheat production per
area
1945-2005 0.9%
Materials Load Bearing Strength to weight ratio 1880-1980 1.6%
Magnetic Magnetic strength 1930-1980 6.1%
Magnetic coercivity 8.1%
Sources, from top: (Nordhaus,1997; Azevedo, 2009; Sheats et al, 1996; Lee, 2005;
Martinson, 2007; Suzuki, 2010; Nemet, 2006; Koh and Magee, 2008; Shaw and Seidler, 2001;
Dong et al, 2010; U.S. Department of Agriculture, 2012, NAS/NRC, 1989)
The benefits from increases in physical scale are often confused with economies of scale,
154
particularly with production equipment. Chemical plants, aluminum smelters, and other
material processing plants exhibit economies of scale more than do assembly plants162
because the production equipment for the former benefits more from increases in physical
scale than do the production equipment for assembly plants. For example, with chemical
plants, the costs of pipes vary as a function of radius whereas the outputs from pipes vary as a
function of radius squared. Similarly, the costs of reaction vessels vary as a function of
surface area (radius cubed) whereas the output of a reaction vessel varies as a function of
radius cubed. These advantages of increases in physical scale have been confirmed in
empirical analysis where capital costs of chemical plants rise much slower than does output
as the physical scale of the pipes and reaction vessels are increased163. Contrast this with
assembly equipment in which physical scale is rarely increased and instead labor is merely
replaced with machines.
Table 7.2. Different Classes of Materials that were found for Each Technology
Technology
Domain
Sub-
Technolog
y
Dimensions
of measure
Different Classes of Materials
Energy
Trans-
formation
Lighting Light
intensity per
unit cost
Candle wax, gas, carbon and tungsten filaments,
semiconductor and organic materials for LEDs
LEDs Luminosity
per Watt
Group III-V, IV-IV, and II-VI semiconductors
Organic
LEDs
Small molecules, polymers, and
phosphorescent materials
Solar Cells Power output
per unit cost
Silicon, Gallium Arsenide, Cadmium Telluride,
Cadmium Indium Gallium Selenide, Dye-
Sensitized, Organic
Energy
storage
Batteries Energy stored
per unit
Lead acid, Nickel Cadmium, Nickel Metal
Hydride, Lithium Polymer, Lithium-ion
Capacitors Carbons, polymers, metal oxides, ruthenium
155
volume, mass
or cost
oxide, ionic liquids
Flywheels Stone, steel, glass, carbon fibers
Information
Trans-
formation
Organic
Transistors
Mobility
(cm2/ Volt-
seconds)
Polythiophenes, thiophene oligomers, polymers,
hthalocyanines, heteroacenes,
tetrathiafulvalenes, perylene diimides
naphthalene diimides, acenes, C60
Living
Organisms
Biological
transfor-
mation
U.S. corn
output per
area
Open pollinated, double cross, single cross,
biotech GMO
Materials Load
Bearing
Strength to
weight ratio
Iron, Steel, Composites, Carbon Fibers
Magnetic Strength Steel/Alnico Alloys, Fine particles, Rare earths
Coercivity Steel/Alnico Alloys, SmCo, PtCo, MaBi,
Ferrites,
Sources, from top to bottom: (Azevedo, 2009; Sheats et al, 1996; Lee, 2005; U.S. DOE,
2010; Tarascon 2009; Naoi and Simon, 2008; Bolund et al, 2007; Shaw and Seidler, 2001;
Dong et al, 2010; Troyer, 2006: NAS/NRC, 1989;
A more recent technology that benefits from increases in the physical scale of production
equipment is liquid crystal displays. Analyses have found that the capital cost per area output
(substrate area per hour) for one type of LCD manufacturing equipment fell by almost 50%
as the substrate size (and its associated equipment) was increased from 0.17 (so-called
Generation II) to 2.7 square meters (Generation VI)164. Second, the capital cost per area
output, this time for a complete facility, dropped by 36% as the substrate size was increased
from 1.4 (in Generation V) to 5.3 square meters in Generation VIII165 (See Figure 7.1).
Generation XI panels are now being produced in sizes of 10.5 square meters. Third, the most
important material in LCDs, glass, also benefits from increases in the scale of their
production equipment166. Fourth, the benefits from larger substrate size can also be seen in the
156
fact that the average selling price of LCD televisions per meter squared dropped 18.8% a year
from the first quarter of 2003 to the first quarter of 2007167. All of this data suggests that LCD
substrate size and its production equipment greatly benefit from increases in scale and this
has implications for solar cells (see below) and roll-to roll printing.
Benefits from increases in scale also exist outside of production equipment and these
examples have little to do with the traditional notion of economies of scale as seen with
production equipment168. Steam engines, internal combustion engines, electricity generating
plants, chemical factories, aluminum smelters, and many types of transportation equipment
also benefited from increases in scale169. Steam engines and internal combustion engines did
so because the output from a piston and cylinder rise with volume while their costs rise with
the cylinder’s surface area and because larger engines can obtain higher temperatures,
pressures, and thus efficiencies. For example, the price per output of a 20 horsepower steam
engine in 1800 was 1/3 that of a 2 hp engine while the price per output of a 225 HP marine
engine was 26% that of a 2.3 HP engine. If these relationships between scale and price per
output were to hold for the largest steam and marine engines, this would mean that the largest
steam and internal combustion engines would be less than 1% the price per HP of the
smallest engines170.
157
0.1 1 10 100 1000 100001
10
100
1000
10000
Rel
ativ
e P
rice
per
Ou
tpu
tFigure 7.1 Scale vs. Relative Price Per Output
Steam Engine(Horse Power)
Marine Engine(Horse Power)
Chemical Plant: (1000s of tons of ethyleneper year)
Commercial air-craft
(Number of pas-sengers)
Oil Tanker:(1000s of
tons)
Output (Scale)
LCD Mfg Equip: (Panel size)
Aluminum(Cell size in1000s of amps)
Electric PowerPlants (in MW)
The modern day equivalent of the steam engine is the steam turbine and the fact that these
turbines (along with boilers and transmission equipment) benefit from increases in scale is a
major reason that the cost per installed electrical generating capacity dropped from about
$78/kilo-watt for a 100-MW coal-fired plant to about $32/kilo-watt for a 600 MW plant, both
in 1929 dollars (See Figure 7/1) and nuclear plants experienced similar levels of cost
reductions as their scale was increased171. The falling capital costs per output, which was
probably even larger for the increases in scale from single digit MW plants to 100-MW
plants, caused the price of electricity to fall from $4.50 per kilowatt hour in 1892 to about
$0.09 by 1969 in constant dollars172.
Transportation equipment such as oil tankers, freighters, industrial trucks, buses, trains, and
to some extent aircraft also benefit from increases in scale. In addition to the benefits from
158
increasing the scale of engines, the reason is that the cost of the transportation equipment is
largely a function of their outer surface area (e.g., dimension squared) while the output is a
function of volume (dimension cubed). For example, the price per capacity of a 265,000 ton
oil tanker is 59% cheaper than the price of a 38,500 oil tanker, the price of a 170,000 ton
freight vessel is 50% cheaper than a 40,000 ton freight vessel, and the price per passenger
capacity of an 853 passenger A380 is 14% cheaper than that of a 132 passenger A318173. Not
only are labor and fuel costs also less on a per capacity basis for the larger oil tankers, freight
vessels and aircraft than are the smaller ones, the advantages of scale become even more
apparent when one considers that some of the first oil tankers were very small (e.g., 1807
tons in late 19th century) and the first commercial aircraft, the DC-1 (early 1930s), could only
carry 12 passengers. Extrapolating to these extremes suggests that today’s largest oil tanker is
almost 1/20 the price per ton of an 1807-ton tanker and that the A380 has a price per
passenger almost 1/2 that of the DC-1174. In combination with containerized shipping and
better information technology (described below), increases in scale is one of the reasons why
the price of airline tickets has dropped dramatically over the last few decades, the price per
ton-mile of rail freight in the U.S. dropped by 88% between 1890 and 2000, the share of U.S.
GDP for transport dropped by more than 50% between 1860 and 2000 (and by more than
75% if airline travel is not considered)175, and both global trade and global travel grew faster
than did overall economic output in the 20th century.
7.4 Geometrical Scaling – reductions in scale
The concept of geometric scaling also helps us understand when technologies benefit from
reductions in scale. These benefits have almost nothing to do with economies of scale and
more to do with the physical laws that govern a technology and the technology’s geometry.
Reducing the scale of transistors, storage regions, and other dimensional features has led to
159
many orders of magnitude improvements in the cost and performance of ICs, magnetic and
optical discs, and newer types of ICs such as MEMS and bio-electronic ICs (see Table 7.3).
This is because for these technologies, reductions in scale lead to improvements in both
performance and cost. For example, placing more transistors or memory cells in a certain area
of an IC increases the speed and functionality and reduces both the power consumption and
size of the final product, which are typically considered improvements in performance for
most electronic products; these reductions in scale also lead to lower material, equipment,
and transportation costs. The combination of both increased performance and reduced costs
as size is reduced has led to many orders of magnitude improvements in the performance to
cost ratio of many ICs. For example, three orders of magnitude reductions in transistor length
have led to about nine orders of magnitude improvements in both the cost of an individual
transistor and the number of transistors on a chip176. Similar arguments can be made for
magnetic and optical storage. Reductions in the magnetic storage area enabled increases in
the magnetic recording density of magnetic cores, drums, disks, and tape, which led to
improvements in both speed and cost. For optical discs, reductions in the wavelength of light
emitted by semiconductor lasers are needed to reduce the size of storage cells for them.
Table 7.3 Technologies that Benefit From Reductions in Scale
Technology
Domain
Sub-
Technology
Dimensions of measure Time
Period
Improvement
Rate Per
Year
Information
Trans-
formation
ICs Number of transistors per
chip/die
1960-2005 33%
MEMS Printing Drops per second for ink jet
printer
1985-2009 61%
Information Magnetic Tape Bits per unit cost 1955-2004 40%
Bits per unit volume 1955-2004 10%
160
Storage
Magnetic Disk Bits per unit cost 1957-2004 39%
Bits per unit volume 1957-2004 33%
Optical Disk Bits per unit cost 1996-2004 40%
Bits per unit volume 1996-2004 28%
Living
Organisms
Biological
transformation
Genome sequencing per unit
cost
1965-2005 35%
MEMS: micro-electronic mechanical systems; LEDs: Light Emitting Diodes; ICs: Integrated
Circuits; Magnetic Resonant Imaging
Sources, roughly from top: (Moore, 2004; Stasiak et al, 2009, Koh and Magee, 2006;
NHGRI, 2012)
Looking to newer technologies, similar arguments can be made for MEMS, bio-electronic
ICs, and DNA sequencing equipment. MEMS are used in motion sensors for Nintendo’s Wii,
nozzles for ink jet printers, in the sensing for micro-gas analyzers, and in the building blocks
for optical computing (e.g., waveguides, couplers, resonators, and splitters) and they are
fabricated using some of the same equipment and processes that are used to construct ICs.
Reductions in the scale of the relevant dimensions exponentially increase the performance of
some types of MEMS and also the number of transistors available for processing the
information. Bio-electronic ICs are basically a MEMS with micro-fluidic channels and they
are used to sense and analyze biological material in for example point-care diagnostics, to
provide better forms of drug delivery in for example IC-controlled smart pills, and to control
artificial implants177. Although DNA sequencers use a variety of different materials and
processes, all of them involve reductions in the scale of the relevant features and these
reductions in physical scale area the major reason for the multiple orders of magnitude
161
reductions in the cost of sequencing and synthesizing DNA178.
Of course, benefiting from either reductions or increases in scale is not a simple task and
they often depend on advances in science and improvements in components. For example,
advances in solid state, plasma, and quantum physics were needed for reducing the scale of
transistors and advances in giant magneto resistance were needed to reduce the scale of
magnetic storage regions on magnetic disks or tape. Similar advances in science were needed
for increasing the scale of the technologies in the previous section. This includes
thermodynamics and heat transfer for engines, aerodynamics for aircraft, and material science
for all of these technologies. As for components, achieving reductions in the size of
transistors required better manufacturing equipment and many of these equipment (e.g.,
photolithographic, plasma etchers) required advances in science and also benefited from
increases in their scale and those of wafers (like LCDs). Larger engines and transportation
equipment required better materials and components with finer tolerances. These finer
tolerances required improvements in manufacturing equipment which in turn required both
better components and manufacturing equipment, along with advances in science179.
Similarly, larger aircraft has required improvements in aluminum, jet engines, and more
recently composites; the weight of aircraft has dropped significantly over the last 20 years as
the strength to weight ratio of materials has been increased several times180.
Some readers might call geometrical scaling and the activities associated with its
implementation “learning” since all improvements involve some form of learning and a
certain type of learning is probably required to exploit geometrical scaling. However, so-
called learning or experience curves focus on cumulative production and imply that learning
done outside of the factory is either unimportant or is being driven by the production of the
final product. Furthermore, since much of the management literature on learning primarily
focuses on the organizational processes that are involved with learning, this literature implies
162
that the organization and not the characteristics of the technology is the bottleneck for
improving the performance or costs of a technology181. Thus, while the management literature
on learning implies that solving energy and environmental problems is primarily an
organizational issue, the concept of geometrical scaling reminds us that the potential for
improving the cost and performance of a technology depends on the characteristics of the
technology182. Without a potential for improvements through for example the two
mechanisms addressed in this paper, it would be difficult for organizational learning to have a
large impact on the costs and performance of a technology no matter how innovative is the
organization.
7.5 Impact on Higher-Level Systems
The largest benefit from the two mechanisms covered above comes from the impact of
specific components on higher level systems. The rapid rates of improvements that were
described above for ICs, magnetic tape and discs, optical discs, liquid crystal displays
(LCDs), and other electronic components have had a large impact on higher-level systems
such as computers, telecommunications equipment, televisions, mobile phones, and DNA
sequencing machines. They did so because these components have large impact on the
performance of these systems and because they represent a large percentage of these costs.
For example, cost breakdowns of electronic products such as computers, mobile phones,
game consoles, set-top boxes, and eReaders have found that 95% of their system costs are
represented by the cost of components and thus increases in cumulative production and
learning in assembly can only impact on about 5% of the system costs. Furthermore, since
many of the assembly operations are standard ones that are used across many different
electronic products, learning in factories probably has a small impact even on the 5% of the
costs that are represented by the assembly operations. This is one reason why the assembly of
163
most electronic products is outsourced to other firms such as Foxconn. When the assembly
operations require little skills or learning, these operations are easily outsourced to other
firms.
Instead, it is changes in the product design in response to the improvements in components
that lead to improvements in the system. For example, rapid improvements in the recording
density of magnetic disk platters or tape did not merely lead to improvements in the first
magnetic disks or reel-to reel tape; they also led to changes in the way we design magnetic
storage systems or what scholars call technological discontinuities (See Table 7.4).
Technological discontinuities involve changes in either the concepts or architectures (i.e.,
linkages between components) that form the basis for a product183. For example, magnetic
storage on disks is based on a concept different from that of tape while changes in the size of
a hard disk drive or magnetic tape system involve changes in the architecture. Smaller disk
drives and tape players emerged and diffused as increases in the magnetic recording density
of platters and disks made these smaller ones economically feasible184.
Similar stories can be told for computers, mobile phones, and other electronic products.
Improvements in electronic components (and also magnetic storage) did not merely lead to
improvements in the first kind of computers, mainframe computers; they also led to the
introduction of scaled-down computers that involve small yet significant changes in the
computer’s architecture. Computers only benefit from some increases in scale and instead
smaller ones with different architectures have provided advantages over mainframe
computers where these advantages are now demanded by most users. This includes easier
customization (first with mini- computers), faster response time (first with personal
computers or so-called PCs)185 and different levels of portability in laptops, PDAs (personal
digital assistants), and tablet computers.
For mobile phones, improvements in electronic components also did not merely lead to
164
improvements in for example the first private mobile radio systems that were introduced in
the 1920s; these improvements led to changes in the way that mobile phone systems are
designed. For example, improvements in ICs led to improvements in switching systems,
which were necessary to enable the automatic switching of users between cells and thus the
implementation of “cellular” phone systems in the early 1980s. Further improvements in ICs
led to changes from analog to digital and to so-called third (and soon fourth) generation
systems in which these improved ICs enabled the use of more sophisticated algorithms for
transmitting digital signals between base stations and phones and that enabled more calls or
data transfers in a specific frequency band. Now improvements in ICs and LCDs are driving
improvements in smart phones and thus the access of Internet content on phones.
Table 7.4 Components, Systems, and Discontinuities
Components Systems Examples of Discontinuities
Integrated circuits Computers Mini-computers
Personal computers
Laptop computers
Personal digital assistants
Mobile phones Analog
Digital
Third Generation systems
Mobile Internet
Consumer electronics Transistor-based devices
Digital devices
Semiconductor
Manufacturing
Equipment
Semiconductors Integrated Circuits
Microprocessors
Memory
Application Specific Integrated Circuits
Application Specific Standard Products
Displays Active-matrix liquid crystal displays
Magnetic tape Music recording and
playback
Reel-to reel tape
8-Track tape
165
Cassette tapes
Video recording and
playback
Quadruplex
Helical Scan (e.g., VHS)
Video cameras
Hard disk drive
platters
Hard Disk Drives Many smaller ones (8”, 5.25”, 3.5”,
2.5”, 1.8” and 1”)
Glass fiber,
semiconductor laser,
light amplifiers
Music recording and
playback
Compact discs
Video recording and
playback
First and Second Generation Digital
Video Disks (DVD)
Telecommunication Fiber optics
All optical systems
Computers, fiber
optics, hard disk
drives
Internet Music
Video
Web 2.0
Source: Author’s analysis
Interestingly, most of Clayton Christensen’s so-called disruptive innovations benefited
from exponential improvements in a “component” and sometimes in a “system.” This
includes disk drives, computers, tape players, and higher-level systems such as retail, health
care, and education186. Disruptive innovations are those innovations that begin from the low-
end of the market and that displace the mainstream technologies. While Clayton Christensen,
who many consider to be the guru of innovation187, primarily focuses on finding a low-end
innovation188, all low-end innovations do not displace the mainstream technologies and thus
become disruptive innovations. This book’s analysis suggests that one can find both
disruptive innovations and high-end technological discontinuities that displace the
mainstream technology by focusing on technologies that are experiencing exponential
improvements in cost and performance.
7.6 Discussion
166
This chapter presents a reality of technology change that is very different from the ones
described in the existing literature. This large difference is why many of the chapters focused
on disproving myths, myths that distort our reality of technology change and thus prevent
firms from introducing effective strategies, governments from introducing effective policies,
and universities from helping students address sustainability. Some of these myths are related
to demand and these include S-curves, linked S-curves, and cumulative production drives
cost reductions. Other myths over-emphasize factory floor activities such as in factory
learning and these include product design enables performance increases and process design
changes enable cost reductions and cumulative production drives cost reductions.
Together the myths associated with demand and factory floor activities make it easy to
think that all technologies have the same potential for rapid improvements. If demand and
commercial production begin, costs will fall and performance will rise. Thus, firms and
governments must stimulate demand in order to move down the learning curve.
This is one problem with Clayton Christensen’s theory of disruptive innovation. By
ignoring rates of improvement and focusing readers on a graph showing how low-end
innovations replace the dominant technologies, his theory uses all of the myths to convince
readers that disruptive technologies are everywhere. Once a niche is found, wallah, the new
technology disrupts the existing one.
As is probably clear by now, this book concludes that the supply part of technology change
is much more important than the demand part of technology change. R&D done by
universities, government laboratories and large firms gradually enables some technologies to
become economically feasible and identifying these technologies is an important challenge
for R&D managers. The rates of improvement can help in these choices along with the
assessments by managers concerning the potential for creating new materials or the benefits
from making changes in physical scale, particularly reductions in scale.
167
For firms, new systems are often where large amounts of value are created. New computers,
mobile phones, the Internet, and all of the new services that have been built on these new
systems have created vast amounts of value. As long as the improvements in electronic
components continue, new systems will continue to emerge and finding these new systems
should be the focus of firms, governments, and universities.
This also has a large impact on public infrastructure. Improvements in electronic systems
are impacting on transportation, health care, education, and other services that have
traditionally been difficult to improve. This has implications for government funding as
governments attempt to balance funding for these subsidized industries with funding for the
technologies that will disrupt them.
8. References
Abernathy W and Utterback J, 1975. Patterns of technological innovation, Technology
Review 80, June/July 1978, pp. 40-47.
Adler P and Clark K 1991. Behind the Learning Curve: A Sketch of the Learning Process,
Management Science 37(3): 267-281.
Adner R 2002. When are Technologies Disruptive? A Demand-Based View of the Emergence
of Competition, Strategic Management Journal 23: 667-688.
Adner R 2004. A Demand-Based Perspective on Technology Life Cycles, Advances in
Strategic Management 21:25-43.
Adner R and Kapoor R 2010. Value Creation in Innovation Ecosystems: How the Structure of
Technological Interdependence affects Firm Performance in New Technology Generations,
Strategic Management Journal 31: 306-333.
Adner R and D Levinthal, 2001. Demand Heterogeneity and Technology Evolution:
168
Implications for Product and Process Innovation, Management Science 47(5): 611-
628.Argyres N.S. and Silverman, B.S. 2004. R&D organization structure and development
of corporate technological knowledge. Strategic Management Journal 25, 929-958.
Adner R and Zemsky P 2006. A Demand-Based Perspective on Sustainable Competitive
Advantage, Strategic Management Journal 27:215-239.
Afuah A and Utterback J 1997. Responding to Structural Industry Changes: A Technological
Evolution Perspective, Industrial and corporate change 6(1): 183-202
Agarwal, R. and Bayus, B. L. 2002. The market evolution and sales takeoff of product
innovations, Management Science, 48(8), 1024-1041.
Agarwal, R. and Gort, M., 1996, ‘The evolution of markets and entry, exit, and survival of
firms,’ The Review of Economics and Statistics, 78(3), 489-498
Amaya, M and Magee, C. The Progress in Wireless Data Transport and its Role in the
Evolving Internet, MIT Technical Report.
web.mit.edu/~cmagee/www/.../24-theprogressinwirelessdatatransportc.pdf
Anderson B and Jacobsson S 2000. Monitoring and assessing technology choice: the case of
solar cells, Energy Policy 28(14): 1037-1049.
Anscombe N 2005. Photonics Spectra, July 2005, Quantum Dots: Small Structures Poised to
Break Big, http://photonics.com/Article.aspx?AID=22350
Argote L 1999. Organizational Learning: Creating, Retaining and Transferring Knowledge,
NY: Springer.
Argote L and Epple D 1990. Learning Curves in Manufacturing, Science 247(4945): 920-
924.
Arrow K 1962. The economic implications of learning by doing, The review of economic
Studies 29(3): 155-173.
Arthur B 2007. The structure of invention, Research Policy 36(2): 274-287.
Axelrod L., Daze R. and Wickham, H. 1968. The large plant concept, Chemical Engineering
169
Progress 64(7): 17
Ayres, R. 1992. Experience and the life cycle, Technovation 12(7): 465-486
Azevedo I, Morgan G, Morgan F, 2009. The Transition to Solid State Lighting, Proceedings
of the IEEE 97: 481-510.
Bae W, Kwak J, Lim J, Lee D, Nam, M. Char, K, Lee C, Lee S. 2010. Multicolored Light-
Emitting Diodes Based on All-Quantum-Dot Multilayer Films Using Layer-by-Layer
Assembly Method. Nano Letters 10: 2368-2373.
Balasubramanian N and Lieberman M 2010. Industry Learning Environments and the
Heterogeneity of Firm Performance, Strategic Management Journal 31: 390-412.
Balconi, M, Brusoni S, and Orsenigo, L 2010. In defence of the linear model: An essay,
Research Policy 39(1): 1-13.
Baldwin C and Clark K 2000. Design Rules: the power of modularity, Cambridge, MA: MIT
Press
Basalla G 1988. The Evolution of Technology, Cambridge University Press.
BCC, 2012.http://www.bccresearch.com/market-research/nanotechnology/carbon-nantubes-
markets-technologies-nan024e.html
Benkard L 2000. Learning and Forgetting: The Dynamics of Aircraft Production, American
Economic Review 90(4): 1034–1054.
Birkinshaw J and Lingblad M 2005. Intrafirm Competition and Charter Evolution in the
Multibusiness Firm, Organization Science 16(6): 674-686.
Bolund B, Bernhoff H and Leijon M 2007. Flywheel energy and power storage systems,
Renewable and Sustainable Energy Reviews 11: 235-258
Bonanno, G., Haworth, B., 1998. Intensity of competition and the choice between product
and
process innovation. International Journal of Industrial Organization 16, 495–510.
Bresnahan T and Trajtenberg M 1995. General purpose technologies 'Engines of growth?,
Journal of Econometrics 65(1) 83-108
170
Bright A 1949. The Electric-Lamp Industry. MacMillan. Pages 221–223 describe Moore
tubes.
Brown R 2011. Staying with Copper in a Fiber World, Rural Telecom, March-April: 22-26.
Brown S and Eisenhardt K 1997. The art of continuous change: Linking complexity theory
and time-paced evolution in relentlessly shifting organizations, Administrative science
quarterly: 42: 1-34
Brynjolfsson E and McAfee A 2014. The Second Machine Age: Work, Progress, and
Prosperity in a Time of Brilliant Technologies, NY: Norton.
Butler J 1988. Theories of Technological Innovation as Useful Tools for Corporate Strategy,
Strategic Management Journal 9: 15-29.
Carlson R 2013. http://www.synthesis.cc/cgi-bin/mt/mt-search.cgi?blog_id=1&tag=
Carlson%20Curves&limit=20. Last accessed March 10, 2014
CCAS, 2014. http://www.ccas-web.org/superconductivity. Last accessed February 4, 2014.
Chader G, Weiland J, Humayun M 2009. Artificial Vision: needs, functioning, and testing of
a
retinal electronic prosthesis, Progress in Brain Research 175: 317-331
Chandler, A 1994, Scale and Scope: The Dynamics of Industrial Capitalism, Boston:
Belknap.
Chandler D 2013. New solar-cell design based on dots and wires, MIT News, March 24,
2013.
ChemTech, 2008. Instant insight: Organic field-effect transistors, Highlights in Chemical
Technology, March 18. http://www.rsc.org/Publishing/ChemTech/Volume/2008/04/
organic_transistors.asp
Christensen, C 1992. Exploring the Limits of the Technology S-curve. Part I: component
technologies, Production and operations management 1(4): 334-357.
Christensen, C 1997. The innovator’s dilemma, Harvard Business School Press, Boston, MA
Christensen C and Bower J 1996. Customer power, strategic investment, and the failure of
171
leading firms, Strategic Management Journal 17: 197-218.
Christensen, C, Craig, & Hart, S 2001, The Great Disruption, Foreign Affairs 80(2): 80 - 95.
Christensen C and Raynor M 2003. The Innovator's Solution: Creating and Sustaining
Successful Growth, Harvard Business School Books.
Clark, K.B. 1985. The interaction of design hierarchies and market concepts in technological
evolution, Research Policy 14(5) 235–251.
Computer History, 2014. Logic chips http://www.computerhistory.org/semiconductor/
timeline/1965-Moore.html. Last accessed February 6, 2014
Connectus, 2012. http://www.conectus.org/xxmarket.html
Coughlin T 2012. Magnetic Tape Archiving for media and entertainment content, Post: where
technology and talent meet. http://www.postmagazine.com/Post-Blog/2012/
October/Magnetic-tape-archiving-for-media-entertainment-.aspx. Accessed May 2, 2014
Cowan, R 1990, ‘Nuclear Power Reactors: A Study in Technological Lock-in,’ Journal of
Economic History, vol. 50, pp. 541-567.
Danner A 2012. Personal communication, September 3, 2012.
David P 1990 The Dynamo and the Computer: An Historical Perspective on the Modern
Productivity Paradox, American Economic Review 80(2): 355-61
de Figueirdo, J., Kyle, M. 2006, Surviving the gales of creative destruction: the determinants
of product turnover, Strategic Management Journal, 27(3): 241-264
de Weck, O., Roos, D., and Magee, C. 2011. Engineering Systems, Cambridge: MIT Press.
Display,Search,2009. http://www.ecnmag.com/news/2009/04/worldwide-oled-revenues-
forecast-reach-55b-2015
Devoret M and Schoeldopf R 2013. Superconducting Circuits for Quantum Information,
Science
339, March 8: 1169-1174
Display search, 2013 http://www.oled-info.com/displaysearch-most-oled-material-market-
growth-will-come-oled-tvs-starting-2014
172
Dong H, Wang C, Hu W 2010. High Performance Organic Semiconductors for Field-Effect
Transistor, Chemical Communications 46: 5211-5222
Dosi, G. 1982. Technological paradigms and technological trajectories, Research Policy 11
(3): 147-162.
Dosi, G. and R. Nelson, 2010. Technical Change and Industrial Dynamics as Evolutionary
Processes, in B. Hall and N. Rosenberg (eds), Handbook of The Economics of Innovation,
Burlington: Academic Press.
Dowling M and McGee, 1994. Business and technology strategies and new venture
performance: A study of the telecommunications equipment industry, Management Science
40(12): 1663-1677.
Dutton J and Thomas A 1984; Treating Progress Functions as a Managerial Opportunity,
Academy of Management Review 9(2): 235-247.
D-Wave, 2013. http://www.dwavesys.com/en/dev-tutorial-hardware.html
Economist, 2012. Television Making: Cracking Up, January 21, 2012, p. 66
Economist, 2012b. http://www.economist.com/node/21540385, Dec 3, 2011, Resistance is
Futile
Economist, 2013. http://www.economist.com/news/21566414-alternative-energy-will-no-
longer- be-alternative-sunny-uplands.
EDN, 2014. Electronic Design News. http://www.edn.com/design/consumer/4397517/2/
Eisenhardt, K.M. and Martin, J.A. 2000. Dynamic capabilities: What are they? Strategic
Management Journal 21: 1105-1121.
Eldredge, N and Gould S, 1972. Punctuated equilibria: an alternative to phyletic gradualis, In
Schopf T (ed), Models in Paleobiology. San Francisco: Freeman Cooper, 82-115.
Evans P et al 2011. 1.12 Tb/s superchannel coherent PM-QPSK InP transmitter photonic
integrated circuit (PIC), Optics Express 19(26): B154-B158
Ferain I, Colinge C, Colinge J 2011. Multigate transistors as the future of classical metal–
173
oxide–semiconductor field-effect transistors, Nature 479, 310–316.
FeRAM: http://en.wikipedia.org/wiki/Ferroelectric_RAM
Fleming L 2001. Recombinant Uncertainty in Technological Search, Management Science
47(1): 117-132.
Fleming L and Sorenson O 2001. Technology as a complex adaptive system: evidence from
patent data, Research Policy 30: 1019-1039.
Fontana R, Robertson N, Hetzler S 2008. Thin-Film Processing Realities for Tbit/in2
Recording,
IEEE Transactions on Magnetics 44(11): 3617-3620.
Foster, R. 1986. The Attacker’s Advantage, NY: Basic Books.
Francis G 2011. Data Storage – Trends and Directions http://lib.stanford.edu/files/pasig-
jan2012/11B7%20Francis%20PASIG_2011_Francis_final.pdf;
Franklin A 2013. Electronics: The road to carbon nanotube transistors, Nature 498: 443-444
Freeman C 1979. The Determinants of Innovation: Market Demand, Technology, and
the Response to Social Problems, Futures, June: 206-15.
Fujimaki A 2013. superconductivity web21, January 16, 2012.
www.istec.or.jp/web21/pdf/12_Winter/E15.pdf
Funk J 2009. Systems, Components, and Technological Discontinuities: The case of
magnetic recording and playback equipment, Research Policy 38(7): 1079-1216.
Funk J 2013a. Technology Change and the Rise of New Industries, Stanford University Press
Funk J 2013b. What Drives Exponential Improvements? California Management Review,
Spring
Fujimoto T 2013. The Long Tail of the Auto Industry Life Cycle, Journal of Product
Innovation
Management 31(1): 8-16.
FutureFab 2013. http://www.future-fab.com/documents.asp?d_ID=4926
Gaimon C, 2008. The Management of Technology: A Production and Operations
174
Management Perspective, Productions and Operations Management 17(1): 1-11.
Garcia R and Calantone R 2002. A critical look at technological innovation typology and
innovativeness terminology: a literature review, Journal of Product Innovation
Management 19(2): 107-182.
Gersick C 1991. Revolutionary Change Theories: A Multilevel Exploration of the Punctuated
Equilibrium Paradigm, The Academy of Management Review 16(1): 10-36.
Gilder G 2000. Telecosm: How Infinite Bandwidth will Revolutionize Our World, Free Press.
Godin B and Lane J 2013. Pushes and Pulls: The His(story) of the Demand Pull Model of
Innovation, Science, Technology and Human Values 38 (5): 621-54.
Gold, B. 1974. Evaluating Scale Economies: The Case of Japanese Blast Furnaces, The
Journal of Industrial Economics 23(1): 1-18
Gonzalez M, 2010. Embedded Multicore Processing for Mobile Communication Systems,
ruhr-uni-bochum.de/integriertesysteme/emuco/files/hipeac_trends_future.pdf.accessed May 16,
2012.
Gort, M. and Klepper, S., 1982, Time Paths in the Diffusion of Product Innovations,
Economic Journal, 92, 630-653.
Gould S and Eldredge N 1977. Punctuated equilibria: the tempo and mode of evolution
reconsidered, Paleobiology 3 (2): 115-151.
Green M, 2009. The Path to 25% Silicon Solar Cell Efficiency: History of Silicon Cell
Evolution, Progress in Photovoltaics 17: 183-189
Green M, Ho-Baillie A, and Snaith H 2014. The emergence of perovskite solar cells, Nature
Photonics 8: 506-514.
Green, P. and Wind, Y. 1973, Multi-attribute Decisions in Marketing: A Measurement
Approach, Dryden Press, Hinsdale, IL.
Griliches Z 1957. Hybrid Corn: An Exploration in the Economics of Technological Change,
Econometrica 25 (4), 501-522.
Haitz R and Tsao J 2011. Solid State lighting: ‘The case’ 10 years after and future prospects,
175
Phys Status Solidi A 208(1): 17-29.
Hardesty L 2011. Long Live the Qubit, MIT News, 1 June 2011.
Hasegawa T and Takeya J 2009. Organic field-effect transistors using single crystals, Science
and Technology of Advanced Materials 10, July, 1-16.
Hatch N and Mowery D 1998. Process Innovations and Learning by Doing in Semiconductor
Manufacturing, Management Science 44 (11): 1461-1477.
Helfat, C.E. and Raubitschek, R.S. 2000. Product sequencing: Co-evolution of knowledge,
capabilities and products. Strategic Management Journal 21: 961~979.
Helfat, C.E. and Peteraf, M.A. 2003. The dynamic resource-based view: Capability lifecycles.
Strategic Management Journal 24, 997-1010.
Helpman E 2003. General Purpose Technologies and Economic Growth, MIT Press.
Henderson, R. 1995. Of life cycles real and imaginary: The unexpectedly long old age of
optical
lithography, Research Policy 24(4): 631-643.
Henderson, R. and Clark, K., 1990, Architectural Innovation: The Reconfiguration of
Existing Product Technologies and the Failure of Established Firms, Administrative
Science Quarterly 35: 9-30.
Hirsh R 1989. Technology and Transformation in the Electric Utility Industry, Cambridge
University Press.
Horowitza 2011. The organic transistor: state-of-the-art and outlook, European. Physics
Journal
Applied Physics 53 (33602)
Hou J and Guo X, 2013. Active Layer Materials for Organic Solar Cells, in Organic Solar
Cells, Choy W (ed), London: Springer Verlag
Hughes T 1983. Networks of Power, Baltimore: Johns Hopkins Press
Iansiti 1995. Technology integration, Research Policy 24(4): 521-542
IDTE, 2012. http://www.idtechex.com/research/reports/organic-photovoltaics-opv-2012-
176
2022-
technologies-markets-players-000319.asp
Iijima S 1991. Helical microtubules of graphitic carbon, Nature 354: 56-58.
iNEMI, 2010. International Electronics Manufacturing Initiative: Mass Data Storage
Roadmap.
ICKnowledge,2009.www.icknowledge.com/economics/fab_costs.html. last accessed
12/7/2009
Investor, 2013. http://investorshub.advfn.com/Quantum-Materials-Corporation-QTMM
-15185/
IPCC, 2013. Renewable Energy Sources and Climate Change Mitigation: Special Report of
the
Intergovernmental Panel on Climate Change. Cambridge University Press. 2013
ISSCC, 2013. International Solid State Circuits Conference, ISSCC Trends
iSuppli, 2012. https://technology.ihs.com/411502/many-iphone-5-components-change-but-
most-suppliers-remain-the-same-teardown-reveals, last accessed on April 21, 2014.
ITRS (International Technology Roadmap for Semiconductors, 2013
Jones N 2013. The Quantum Company, Nature 498, 20 June, 286-288.
Jorgenson D, Ho M, and Stiroh K 2008. A Retrospective Look at the U.S. Productivity
Growth
Resurgence, Journal of Economic Perspectives 22(1): 3-24.
Kalender W. 2006. X-ray computed tomography, Physics in Medicine and Biology 51: 29-53.
Kapoor R and McGrath P 2014. Unmasking the interplay between technology evolution and
R&D collaboration: evidence from the global semiconductor manufacturing industry, 1990-
2010, Research Policy 43: 555-569.
Klepper, S. 1996. Entry, exit, growth and innovation over the product life cycle, American
Economic Review 86(3) 562–583.
Klepper S 1997. Industry Life Cycles, Industrial and Corporate Change 6(1): 145-81
177
Klepper S and Thompson 2006. Submarkets and the Evolution of Market Structure, The
RAND Journal of Economics 37(4): 861-886.
Klevorick A, Levin R, Nelson R, Winter S 1995, on the sources and significance of
interindustry differences in technological opportunities, research policy 24: 185-205
Koh, H. and Magee, C. 2006. A functional approach for studying technological progress:
Application to information technologies, Technological Forecasting and Social Change
73: 1061-1083.
Koh, H. and Magee, C. 2008. A functional approach for studying technological progress:
Extension to energy technology, Technological Forecasting and Social Change 75:
735-758.
Koomey J, Berard S, Sanchez M, Wong H, 2011. Implications of Historical Trends in the
Electrical Efficiency of Computing, IEEE Annals of the History of Computing 33(3): 46-54.
Kressel H and Lento T, 2007. Competing for the Future: How Digital Innovations are
Changing the World, NY: Cambridge University Press.
Kuhn K 2009. Moore's Law past 32nm: Future Challenges in Device Scaling, 13th
International
Workshop on Computational Electronics.
ftp://download.intel.com/newsroom/bios/pdfs/kkuhn/Kuhn_IWCE_invited_text.pdf
Kurzwell, R., 2005, The Singularity is Near, NY: Penguin Books.
Kwak J 2010. PhD Thesis, cited in Changhee Lee, Seoul National University
www.andrew.cmu.edu/org/nanotechnology-forum/Forum_7/Presentation/CH_Lee.pdf
Kwak J, Bae W, Lee D, Park I, Lim J, Park, M, Cho H, Woo H, Yoon D, Char K, Lee S, Lee
C
2012. Bright and Efficient Full-Color Colloidal Quantum Dot Light-Emitting Diodes Using
an
Inverted Device Structure, Nano Letters 12: 2362−2366
Lapre M, Mukherjee A, Wassenhove L 2000. Behind the Learning Curve: Linking Learning
178
Activities to Waste Reduction, Management Science 46(5):597-611.
Lee, C 2005. OLED 1 – Introduction,
http://wenku.baidu.com/view/783fa93283c4bb4cf7ecd196.html
Lieberman M 1984. The learning curve and pricing the chemical processing industries. Rand
Journal of Economics 15: 213-228. R&D expenditures and capital intensity
Linton J and Walsh S 2008. Technological Forecasting & Social Change 75: 583–594
Liu, J and Hersam M 2010. Recent developments in carbon nanotube sorting and selective
growth, Materials Research Society Bulletin 35: 315–321
Lipsey, R. Carlaw, K. and Bekar, C. 1998. What Requires Explanation? In General Purpose
Technologies and Economic Growth, Helpman (ed), MIT Press.
Lipsey, R. Carlaw, K. and Bekar, C. 2005. Economic Transformations, NY: Oxford Univ
Press.
Loch C and Huberman B 1999. A Punctuated-Equilibrium Model of Technology Diffusion,
Management Science 45(2): 160-177.
Mannan, S. 2005. Lee’s Loss Prevention in the Process Industries, Vol. 1, Burlington, MA:
Elsevier Butterworth-Heinemann.
Markets and Markets, 2011. http://www.marketsandmarkets.com/Market-
Reports/printed-electronics-market-197.html
Margulies M (and more than 20 co-authors) 2005. Genome sequencing in microfabricated
high-density picolitre reactors, Nature. 2005 September 15; 437(7057): 376–380.
Martino J 1971. Examples of Technological Trend Forecasting for Research and
Development Planning. Technological Forecasting and Social Change 2: 247-260
Martinson R 2007. Industrial markets beckon for high-power diode lasers, Optics, October:
26-27. ww.nlight.net/nlight-files/file/articles/OLE%2010.2007_Industrial%20markets...pdf
Mas-Torrent M and Rovira C 2008. Novel small molecules for organic field-effect transistors:
towards processability and high performance, Chemical. Society Review 37: 827.
179
Mathews J and Cho D 1999. A New Model for Global Growth, Journal of World Business, 34
(2): 139-156.
McGahan A and Silverman B 2001. How does innovative activity change as industries
mature?,
International Journal of Industrial Organization 19: 1141–1160
Mellor C 2012. http://www.theregister.co.uk/2012/05/09/wd_disk_tech_views/. WD bigshots
spin superfast disk roadmap, 9 May 2012, The Register.
Merali Z 2011. First sale for quantum computing, Nature 474(18): 194-198.
Miller G 2012. New power semiconductor technologies challenge assembly and system
setups.
Embedded.com, April 16,
Moore, G.E. 2006, Moore’s law at 40. Chapter 7 in understanding Moore’s law. In: Four
Decades of Innovation; Brock, D.C., Ed.; Chemical Heritage Foundation: Philadelphia, PA.
Mowery D and Rosenberg N (1979), The Influence of Market Demand Upon Innovation: A
Critical Review of Some Recent Empirical Studies, Research Policy 8: 102-53.
MRAM: http://en.wikipedia.org/wiki/Magnetoresistive_Random_Access_Memory
Mukhanov O and Semenov V 1985. A novel way of digital information processing in the
Josephson junction circuits, Preprint No. 9/1985, Dept. of Physics, Moscow State
University.
Mudambi R and Swift T 2013. Knowing when to Leap: Transitioning between Exploitative
and Explorative R&D, Strategic Management Journal 35: 126-145.
Murmann J P and Frenken K 2006. Toward a systematic framework for research on dominant
designs, technological innovations, and industrial change, Research Policy 3: 925–952
Murray F 2002. Innovation as co-evolution of scientific and technological networks:
exploring
tissue engineering, Research Policy 31: 1389-1403.
180
Murray F 2004. The role of academic inventors in entrepreneurial firms: sharing the
laboratory
life, Research Policy 33: 643-659.
Nagy B, Farmer D, Bui Q, Trancik J 2013. Statistical Basis for Predicting Technological
Progress. PLoS ONE 8(2): e52669. doi:10.1371/journal.pone.0052669NREL, 2013.
Naoi K and Simon P, 2008, New Materials and New Configurations for Advanced
Electrochemical Capacitors, The Electrochemical Society Interface, Spring 17(1): 34-37.
Nemet, G. 2006. Beyond the learning curve: factors influencing cost reductions in
Photovoltaics, Energy Policy 34: 3218-3232
Nemet G 2009. Demand-pull, technology-push, and government-led incentives for non-
incremental technical change, Research Policy 38(5): 700-709.
NHGRI, 2013. National Human Genome Research Institute.
www.genome.gov/sequencingcosts/, last accessed on September 10, 2013.
Nordhaus W 2007. Two Centuries of Productivity Growth in Computing, Journal of
Economic
History 67(1): 128-159.
Nordhaus W 2009. The Perils of the Learning Model for Modeling Endogenous
Technological
Change, Cowles Foundation Discussion Paper No. 1685.
NREL, 2013. National Research Energy Lab.
http://en.wikipedia.org/wiki/File:PVeff(rev131204)a.jpg
OLED, 2013. OLED-info.com. http://www.oled-info.com/oled-mobile-phones, last accessed
on 26 November 2013.
Oliner S and Sichel D 2002. Information technology and productivity: where are we now and
where are we going? Economic Review 91: 1-32
Oliner S, Sichel D, Stiroh K 2007. Explaining a Productive Decade, Finance and Economics
Discussion Series 2007-63,
181
Orton J 2005. Semiconductors and the Information Revolution: Magic Crystals that made IT
Happen, Oxford University Press.
Paranthaman P and Izumi T 2004. High-Performance YBCO-Coated Superconductor Wires,
MRS Bulletin, August 533-536.
Pavitt K 1984. Sectoral patterns of technical change: towards a taxonomy and a theory,
Research
Policy 13: 343–373.
Patel, P 2011. Quantum Dots as Solar Cells, MIT Technology Review.
Perez T and De Rose A 2010. Non-Volatile Memory: Emerging Technologies and their
Impacts
on Memory Systems, Technical Report No 60, Pontifica Universidade Catolica do Rio,
www3.pucrs.br/pucrs/files/uni/poa/facin/pos/relatoriostec/tr060.pdf, accessed June 30, 2014
Paranthaman P and Izumi T 2004. High-Performance YBCO-Coated Superconductor Wires,
MRS Bulletin, August 533-536.
Perlin J 2002. From Space to Earth: The Story of Solar Electricity, Harvard University Press.
Physorg, 2013. http://phys.org/news/2013-01-sony-tvs-high-end-quantum-dot.html
Preil M 2012. Optimum Dose for EUV: Technical vs. Economic Drivers, April 26. Future Fab
41
ProStat Services, 2013.
http://www.prostatservices.com/?p=73&option=com_wordpress&Itemid=56
Research & Markets, 2013. http://www.prnewswire.com/news-releases/quantum-dot-and-
quantum-dot-display-qled-market-shares-strategies-and-forecasts-worldwide-
nanotechnology-
2013-to-2019-212987981.html
Rosenberg, N. 1969. The Direction of Technological Change: Inducement Mechanisms and
Focusing Devices, Economic Development and Cultural Change 18 (1): 1-24.
Rosenberg N 1974. Science, Invention, and Economic Growth, Economic Journal 84: 90-
182
108.
Rosenberg N 1982. Inside the Black Box: Technology and Economics, Cambridge University
Press.
Rosenberg N 1994. Exploring the black box, Cambridge University Press.
Rosenberg N and Mowery D 1988. Paths of Innovation: Technological Change in 20th-
Century
America, Cambridge University Press.
Rosenkranz S 2003. Simultaneous choice of process and product innovation when consumers
have a preference for product variety, Journal of Economic Behavior & Organization 50(2):
183-201.
Ryan D, Rahimi M, Lund J, Mehta R, Parviz BA, 2007. Toward nanoscale genome
sequencing.
Trends in Biotechnology 25(9):385-9.
Saha S 2007. Consumer preferences and product and process R&D, Rand Journal of
Economics 38(1): 250-268.
Sahal D 1985. Technological guideposts and innovation avenues. Research Policy 14(2): 61.
Sanchez R and Mahoney J 1996. Modularity, flexibility, and knowledge management in
product and organization design, Strategic Management Journal 17(52): 63-76.
Sanderson, 2009. Quantum dots go large, Nature 459, 760–761 (10 June 2009)
Sanvido M, Chu F, Kulkarni A, Selinger R 2008. NAND Flash Memory and its Role in
Storage
Architectures, Proceedings of the IEEE, November: 1864-1874.
Schmookler, J. 1966. Invention and Economic Growth, Cambridge, Harvard University Press.
Schumpeter J 1942. Capitalism, Socialism and Democracy. London: Routledge.
Selvamanickam V 2011. Second-generation HTS Wire for Wind Energy Applications,
Symposium on Superconducting Devices for Wind Energy, 25 February, Barcelona, Spain
Shaw J and Seidler P 2001. Organic electronics: Introduction, IBM Journal of Research and
183
Development 45 (1): 3 - 9.
Sheats, J, H Antoniadis, M Hueschen, W Leonard, J Miller, R Moon, D Roitman, A
Stockinget, Organic Electroluminescent Devices, Science 273 (1996): 884-888.
Shiohara T, Yoshizumi M, Takagi Y, Izumi T 2013. Future prospects of high Tc
superconductors-coated conductors and their applications, Physica C: Superconductivity 484,
15 January 2013: 1-5.
Sinclair G, Klepper S, and Cohen W 2000. What’s experience got to do with it? Sources of
cost
reduction in a large specialty chemicals producer. Management Science, 46: 28–45.
Singualrity.com http://singularity.com/charts/ Last accessed on February 27, 2014
SingularityHub.com, 2013. http://singularityhub.com/2012/09/17/new-software-makes
-synthesizing-dna-as-easy-as-using-an-ipad/, last accessed on November 27, 2013.
Smith R 1988. A Historical Overview of Computer Architecture, IEEE Annals of the
History of Computing 10(4): 277-303.
Solow R. 1957. Technical Change and the Aggregate Production Function, Review of
Economics and Statistics, 39: 312-320.
Sood A and Tellis F 2005. Technological Evolution and Radical Innovation, Journal of
Marketing 69: 152–168.
Sood A, James G, Tellis G. and Zhu J 2012. Predicting the Path of Technological Innovation:
SAW vs. Moore, Bass, Gompertz, and Kryder. Marketing Science 31(6): 964-979.
Sorensen J and Stuart T 2000. Aging, Obsolescence and Organizational Innovation,
Administrative Science Quarterly 45(1): 81-112
Stasiak J, Richards S, and Angelos S 2009. Hewlett Packard's Inkjet MEMS Technology,
Proc. of Society of SPIE: 7318, http://144.206.159.178/ft/CONF/16431771/16431793.pdf
Stieglitz N and Heine K. 20007. Innovations and the Role of Complementarities in a Strategic
Theory of the Firm, Strategic Management Journal 28:1-15.
Stobaugh, R 1988. Innovation and Competition: The Global Management of Petrochemical
184
Products. Boston: Harvard Business School Press
Suarez F and Lanzolla G 2007. The Role of Environmental Dynamics in Building a First
Mover Advantage Theory, Academy of Management Review 33(2): 377-399.
Suzuki T, 2010, Challenges of Image-Sensor Development, International Solid State Sensors
Conference.
Tang, C. W.; Vanslyke, S. A. (1987). "Organic electroluminescent diodes". Applied Physics
Letters 51 (12): 913.
Tarascon, J. 2009. Batteries for Transportation Now and In the Future, presented at Energy
2050,
Stockholm, Sweden, October 19-20.
Teece, D.J. 2007. Explicating dynamic capabilities: The nature and microfoundations of
sustainable enterprise performance. Strategic Management Journal 28, 1319-1350.
Thompson P 2012. The Relationship between Unit Cost and Cumulative Quantity and the
Evidence for Organizational Learning-by-Doing, Journal of Economic Perspectives 26(3):
203-224.
Thornton R and Thompson P 2001. Learning from Experience and Learning from Others,
American Economic Review 91(5): 1350-1368.
Tilton J 1971. International Diffusion of Technology: The case of semiconductors,
Washington DC: The Brookings Institution.
Tripsas M 2007. Customer Preference Discontinuities: A Trigger for Radical Technological
Change, Managerial and Decision Economics 29(2-3): 79-97.
Tschang T 2007. Balancing the Tensions between Rationalization and Creativity,
Organization Science 18(6): 989-1005.
Tushman, M. and Anderson, P. 1986. Technological Discontinuities and Organizational
Environment, Administrative Science Quarterly 31: 439-456.
Tushman, M.L. and O'Reilly, C.A. 1996. Ambidextrous organizations: Managing
185
evolutionary and revolutionary change. California Management Review 38(4): 8-31.
Tushman M. and Romanelli E 1985. Organizational evolution: A metamorphosis model of
convergence and reorientation. Research in Organizational Behavior 7: 171-222.Ulrich K
1995. The role of product architecture in the manufacturing firm, Research Policy 24: 419-
440.
Ulrich K and Eppinger S 2011. Product Design and Development, NY: McGraw-Hill.
Ulrich K 2013. Design: Creation of Artifacts in Society, University of Pennsylvania.
Utterback, J., 1994, Mastering the dynamics of innovation, Harvard Business School Press.
U.S. DoE (Department of Energy), 2010. $1/W Photovoltaic Systems: White Paper to
Explore a Grand Challenge for Electricity from Solar
Utterback, J., 1994, Mastering the dynamics of innovation, Harvard Business School Press.
Utterback J and Suarez. 1993. Innovation, competition, and industry structure, Research
Policy 22: 1–21.
van den Ende J, Jaspers F, and Rijsdijk S 2013. Should System Firms Develop
Complementary
Products? A Dynamic Model and an Empirical Test, Journal of Product Innovation
Management 30(6): 1178-1198.
Victor R and Irina K 2000. Electron and photon effects in imaging devices utilizing quantum
dot infrared photodetectors and light emitting diodes. Proceedings of SPIE 3948: 206–219
Walsh V 1984. Invention and Innovation in the Chemical Industry: Demand-Pull and
Discovery-Push? Research Policy 13 (4): 211-34
Weiss M, Junginger M, Patel M, Blok K 2010. A review of experience curve analyses for
energy
demand technologies, Technology Forecasting and Social Change 77: 411-428.
Wheelwright S and Clark 1992. Revolutionizing Product Development: Quantum Leaps in
Speed, Efficiency, and Quality, NY: The Free Press.
Wikipedia,2014. http://en.wikipedia.org/wiki/File:Transistor_Count_and_Moore%27s_Law_-
186
_2011.svg. Last accessed on February 4, 2014.
Winter S 2008. Scaling heuristics shape technology! Should economic theory take notice?
Industrial and Corporate Change 17(3): 513–531.
Wright T P, 1936. Factors Affecting the Cost of Airplane, Journal of Aeronautical Sciences,
3(4): 122 – 128.
Wouters D 2012. Resistive switching materials and devices for future memory applications,
43rd IEEE Semiconductor Interface Specialists Conference, Tutorial SISC San Diego,
December 5
Yole, 2013a. http://www.siliconsemiconductor.net/article/76878-Yole-Non-volatile-
memories-
to-penetrate-many-markets.php
Yole, 2013b. Emerging NVM enter niche memory markets: expected to reach $2B by 2018.
Yoon Y 2010. Nano-Tribology of Discrete Track Recording Media, Unpublished PhD
Dissertation, University of California, San Diego
187
188
1 McGrath R 1998. Technological discontinuities and media patterns: assessing electric vehicle
batteries, Technovation 18(11): 677-687. Anderson B and Jacobsson S 2000. Monitoring and
assessing technology choice: the case of solar cells, Energy Policy 28(14): 1037-1049. Schilling M
and Esmundo M 2009. Technology S-curves in renewable energy alternatives: Analysis and
implications for industry and government, Energy Policy 37(5): 1767-1781
2 Tushman, M. and Anderson, P. 1986. Technological Discontinuities and Organizational
Environment, Administrative Science Quarterly 31: 439-456.
3 For punctuated equilibrium in evolutionary biology see Eldredge, Niles and S. J. Gould
(1972). "Punctuated equilibria: an alternative to phyletic gradualism" For punctuated equilibrium in
organizations, see Gersick C 1991. Revolutionary Change Theories: A Multilevel Exploration of the
Punctuated
Equilibrium Paradigm, The Academy of Management Review 16(1): 10-36. Miller D and Friesen P
1984. Organizations: A quantum view. Englewood Cliffs, NJ: Prentice-Hall. Tushman M. and
Romanelli E 1985. Organizational evolution: A metamorphosis model of convergence and
reorientation. Research in Organizational Behavior 7: 171-222. Loch C and Huberman B 1999. A
Punctuated-Equilibrium Model of Technology Diffusion, Management Science 45(2): 160-177.
4 For example, most of the commercial product types for which Sood and Tellis (2005) report step-
wise functions do not show large numbers of data points for the relevant periods of time. In
extensions to this analysis, they develop a step-wise model (Sood and Tellis, 2009) and contrast this
model to S-curves and other models (Sood et al, 2012). In doing so, they focus on products, many
of which initially failed, thus causing large gaps to emerge in the time series before another product
is released. Although they correctly cite these gaps as evidence of a step-and weight strategy for
products, improvements probably continued to be made on the technologies in laboratories even as
the firms did not introduce new products. This is an important methodological limitation of
analyzing the performance of products before a technology has consistent market success. If one is
to look for an acceleration during the early years of a technology, one must include best laboratory
results. After all, a major objective of such an analysis of performance vs. time curves is to
understand whether an acceleration will occur as demand, investment, or something else increases.
5 Griliches Z 1957. Hybrid Corn: An Exploration in the Economics of Technological Change,
Econometrica 25 (4), 501-522. Mansfield E 1968. The Economics of Technological Change, NY:
W.W. Norton & Company Inc.
6 Early and heavily cited papers concerning this theory include: Utterback J and Abernathy W,
1975. A dynamic model of process and product innovation, Omega, 3(December):639-656.
Abernathy W and Utterback J, 1975. Patterns of technological innovation, Technology Review 80,
June/July 1978, pp. 40-47. Utterback, J., 1994, Mastering the dynamics of innovation, Harvard
Business School Press. Adner R and D Levinthal, 2001. Demand Heterogeneity and Technology
Evolution: Implications for Product and Process Innovation, Management Science 47(5): 611-
628. For the impact on a shakeout in the number of firms, see Klepper, S. 1996. Entry, exit,
growth and innovation over the product life cycle, American Economic Review 86(3) 562–583.
Klepper S 1997. Industry Life Cycles, Industrial and Corporate Change 6(1): 145-81
7 Stobaugh, R 1988. Innovation and Competition: The Global Management of Petrochemical
Products. Boston: Harvard Business School Press
Utterback, J., 1994, Mastering the dynamics of innovation, Harvard Business School Press.
8 Wright T P, 1936. Factors Affecting the Cost of Airplane, Journal of Aeronautical Sciences,
3(4): 122 – 128. Argote L 1999. Organizational Learning: Creating, Retaining and Transferring
Knowledge, NY: Springer. Argote L and Epple D 1990. Learning Curves in Manufacturing, Science
247(4945): 920- 924. Ayres, R. 1992. Experience and the life cycle, Technovation 12(7): 465-486
Thornton R and Thompson P 2001. Learning from Experience and Learning from Others,
American Economic Review 91(5): 1350-1368. Nagy B, Farmer D, Bui Q, Trancik J 2013.
Statistical Basis for Predicting Technological Progress. PLoS ONE 8(2): e52669.
doi:10.1371/journal.pone.0052669NREL, 2013.
9 Schmookler, J. 1966. Invention and Economic Growth, Cambridge, Harvard University Press.
Sinclair G, Klepper S, and Cohen W 2000. What’s experience got to do with it? Sources of cost
reduction in a large specialty chemicals producer. Management Science, 46: 28–45.
10 Christensen, C 1997, The innovator’s dilemma, Harvard Business School Press, Boston, MA
11 (Tushman and Anderson, 1986; Butler, 1988; Utterback, 1994; Gaimon, 2008; Schilling and
Esmundo, 2009; Kapoor and McGrath, 2014)
12 (Foster, 1986; Kelley, 2010)
13 (Tushman and Anderson, 1986)
14 Or in the words of two famous scholars, “technology evolves through relatively long periods of
incremental change punctuated by relatively rare innovations that radically improve the state of the
art.” Building from this paper and findings from evolutionary biology (Eldredge and Gould, 1972),
other scholars use the term “punctuated equilibrium” to describe a long-term process in which
stable environments are punctuated by new technologies that provide jumps in performance (and
perhaps organizational change (Gersick, 1991; Miller and Friesen, 1984; Tushman and Romanelli,
1985; Romanelli, 1994; Loch and Huberman, 1999), followed quickly by a slowdown at least until
the next rare innovation emerges
15 Using “degree of research efforts” instead of time, Foster (1986) provides performance vs. effort
numbers for artificial hearts, nylon, polyester, rayon, and cotton of which only the nylon data partly
exhibits the beginning of an S-curve and thus a large jump in performance. Butler (1988) interprets
ups and downs in linear plots of time vs. three measures of performance for air travel (fatalities per
passenger aircraft mile, average seat miles, and average air speed) as successions of S-curves, but
an S-curve is not shown for any of them in the figures. Christensen (1992) analyzes performance vs.
time for both individual firms and industries in his analysis of magnetic recording density for hard
disk platters. His data for individual firms show the beginnings of an S-curve for some firms, but
not for best performance in the industry. Using industry data on performance vs. time for four types
of products, Sood and Tellis (2005) state “the plots suggest a series of step functions, each of which
could approximate an S curve.” This step-wise model is extended to a larger data set, some of which
are exploratory products, in which a lack of improvements is assumed to mean that firms have
“stepped and waited” (Sood et al, 2012). Tripsas’ (2008) analysis of the typesetter industry shows
two S-curves for hot metal typesetter speeds thus providing some of the best evidence of S-curves.
16 (Foster, 1986; Garcia and Clantone, 2002; Utterback, 1994)
17 (Butler, 1988)
18 (Foster, 1986; Butler, 1988; Utterback, 1994)
19 (Kurzweil, 2005, Brynjolfsson, and McAfee, 2014)
20 (Brynjolfsson and McAfee 2014)
21 For example, most of the commercial product types for which Sood and Tellis (2005) report step-
wise functions do not show large numbers of data points for the relevant periods of time. In
extensions to this analysis, they develop a step-wise model (Sood and Tellis, 2009) and contrast this
model to S-curves and other models (Sood et al, 2012). In doing so, they focus on products, many
of which initially failed, thus causing large gaps to emerge in the time series before another product
is released. Although they correctly cite these gaps as evidence of a step-and weight strategy for
products, improvements probably continued to be made on the technologies in laboratories even as
the firms did not introduce new products. This is an important methodological limitation of
analyzing the performance of products before a technology has consistent market success. If one is
to look for an acceleration during the early years of a technology, one must include best laboratory
results. After all, a major objective of such an analysis of performance vs. time curves is to
understand whether an acceleration will occur as demand, investment, or something else increases.
22 (Economist, 2012)
23 (Nemet, 2005)
24 (Foster, 1986; Butler, 1988; Utterback, 1994)
25 (Amaya and Magee, 2008)
26 (Butler, 1988)
27 (Tushman and Anderson, 1986)
28 (Smith, 1988; Bresnahan and Trajtenberg, 1995)
29 (Foster, 1986; Tripsas, 2008)
30 (Christensen, 1992; Sood and Tellis, 2012)
31 (Sood and Tellis, 2012)
32 (Schmookler, 1966)
33 Foster (1985)
34 (Chesbrough, 2005)
35 (Christensen, 1992)
36 (McGrath, 1998; Anderson and Jacobsson, 2000; Schilling and Esmundo, 2009)
37 (Rockart and Dutt, 2014)
38 . (Foster, 1986; Sahal, 1981; Butler, 1988; Utterback, 1994; Garcia and Calantone, 2002; Sood
and Tellis, 2005)
39 We do this by using a technique developed by ProStat Services (2013) that tests for the
differences between slopes.
40 (Bresnahan and Trajtenberg, 1995)
41 (Mukhanov and Semenov, 1985)
42 (Iijima, 1991)
43 (FeRAM, 2014; MRAM, 2014; Perez and De Rose, 2010; Wouters, 2012)
44 (Henderson, 1995; ITRS, 2013)
45 (Jones, 2013).
46 (Brown, 2011).
47 ICs (ISSCC, 2013; Gonzalez, 2010)
48 (Cowan, 1990)
49 (Perlin, 2002; Nemet, 2006).
50 (NREL, 2014)
51 (Nemet, 2006; Economist, 2012)
52 (Shaw and Seidler, 2001; Green et al, 2014).
53 (Bright, 1949; Orton, 2009; Tang and Vanslyke, 1987)
54 (OLED, 2013)
55 (Investor, 2013)
56 (Victor and Irina, 2000)
57 (Abernathy and Utterback, 1978; Foster, 1986; Sahal, 1981; Butler, 1988; Utterback, 1994;
Garcia and Calantone, 2002; Tripsas, 2007)
58 (Abernathy and Utterback, 1978; Foster, 1986; Sahal, 1985; Butler, 1988; Utterback, 1994;
Garcia and Calantone, 2002)
59 (Foster, 1986; Utterback, 1994; Sood and Tellis, 2005)
60 Klevorick et al, 1995; Balconi et al, 2010
61 (Rosenberg and Mowery, 1979)
62 (Mowery and Rosenberg, 1979)
63 (Christensen, 1997)
64 (ProStat Services, 2013)
65 (David, 1989; Bresnahan and Trajtenberg, 1995; Helpman, 2003; Lipsey et al, 2005).
66 A notable exception is the books by Vaclav Smil on energy and sustainability.
67 (Rosenberg, 1974; Klevorick et al, 1995)
68 (Nagy et al, 2013)
69 (Stobaugh, 1988; Pratten, 1971; Pavitt, 1984; Chandler, 1994; Funk, 2013a; Funk, 2013b; Haldi
and Whitcomb, 1967; Axelrod et al, 1968; Rosenberg, 1994; Mannan, 2005)
70 (Rosenberg, 1994; Winter, 2008; Funk, 2013a: Funk, 2013b)
71 (Stobaugh, 1988)
72 (Pratten, 1971; Pavitt, 1984; Chandler, 1994; Funk, 2013a; Funk, 2013b)
73 (Thompson, 2012)
74 (Arrow, 1962; Wright, 1936, Argote and Epple, 1990; Adler and Clark, 1991; Thornton and
Thompson, 2001; Argote, 1999; Lapre et al, 2000; Utterback, 1994; Benkhard, 2000)
75 (Pratten, 1971; Pavitt, 1984; Chandler, 1994; Rosenberg, 1994; Winter, 2008; Funk, 2013a: Funk,
2013b)
76 (Utterback, 1994)
77 (Schmookler, 1966)
78 (Lieberman, 1984; Dutton and Thomas, 1984; Balasubramania and Lieberman, 2010)
79 (Rosenberg, 1974).
80 (Klevorick et al, 1995)
81 Rosenberg, 1974
82 Adner and Levinthal 2001; Bonanno and Haworth, 1988; Butler, 1988; Fujimoto and Clark, 1990;
Afuah and Utterback, 1994; Klepper, 1996; Adner and Levinthal 2001; Adner, 2002, 2004;
Rosenkranz, 2003; Birkinshaw and Lingblad, 2005; Adner and Zemsky, 2006; Saha, 2007; Tschang,
2007; Utterback and Abernathy, 1975; Abernathy and Utterback, 1978; Utterback, 1994; van den
Ende et al, 2013)
83 (Gort and Klepper, 1982; McGahn and Silverman, 2001)
84 Stobaugh, R 1988. Innovation and Competition: The Global Management of Petrochemical
Products. Boston: Harvard Business School Press. Utterback, J., 1994, Mastering the dynamics of
innovation, Harvard Business School Press.
85 (Smith, 1988).
86 (Adner and Levinthal, 2001)
87 (Green and Wind, 1979),
88 (Adner, 2002, 2004; Adner and Zemsky, 2006; de Figueiredo and Kyle, 2006)
89 (US DoE, 2010; Azevedo et al, 2009; Haitz and Tsao, 2011; Koh and Magee, 2006, 2008)
90 (Solow, 1957; Schumpeter, 1942)
91 (Weiss et al, 2010; Nagy et al, 2013)
92 (Stobaugh, 1986; Utterback, 1994)
93 (Utterback, 1994; Adner and Levinthal, 2001
94 (Adner and Levinthal, 2001)
95 (Wright, 1936; Argote, 1999; Arogote and Epple, 1990; Ayres, 1992; Thornton and Thompson,
2001; Nagy, Farmer, Bui, Trancik, 2013)
96 (Christensen, 1997)
97 (Azevedo et al, 2009)
98 (OLED, 2013; Display Search, 2009; Display Search 2013)
99 Sheats et al (1996) and Lee (2005)
100 (Nordhaus, 2007)
101 (Koh and Magee, 2006)
102 (ChemTech, 2008; Mas-Torrent and Rovira, 2008)
103 (Dong et al, 2010)
104 (Horowitza, 2011)
105 (US DoE, 2010)
106 (IDTE, 2012)
107 (Hou and Guo, 2013)
108 (US DoE, 2010)
109 (Investor, 2013)
110 (Research & Markets, 2013)
111 (Patel, 2011; Chandler, 2013)
112 (Azevedo et al, 2009)
113 (Physorg, 2013)
114 (Anscombe, 2005; Sanderson, 2009)
115 (Bae et al, 2010)
116 (Kwak et al, 2012)
117 (Moore, 2006)
118 (Yole, 2013a)
119 (ISSCC, 2012; Yole, 2013b)
120 (Yole, 2013b)
121 (Kuhn, 2009)
122 (BCC, 2012)
123 (Franklin, 2013)
124 Liu and Hersem (2010)
125 (Franklin, 2013)
126 (Franklin, 2013)
127 (CCAS, 2014)
128 (CCAS, 2014)
129 (Economist, 2012b)
130 (Connectus, 2012’ Selvamanickam, 2011)
131 (Shiohara et al, 2013; Paranthaman and Izumi, 2004)
132 (Koh and Magee, 2006; Koomey et al, 2011)
133 (Merali, 2011)
134 (Fujimaki, 2012)
135 (Jones, 2013)
136 (Hardesty, 2011)
137 (Devoret and Schoelkopf, 2013; Hardesty, 2011)
138 (D-Wave, 2013; Jones, 2013)
139 (Hughes, 1984; Rosenberg, 1969; Dosi and Nelson, 2010)
140 (Sinclair et al, 2000)
141 (Liu and Hersem, 2010; Franklin, 2013)
142 Winter (2008) and Funk (2013a and 2013b)
143 (Wright 1936; Arrow 1962; Argote and Epple 1990; Ayres 1992; Nagy et al, 2013)
144 (Weiss et al, 2010).
145 (Henderson and Clark, 1990) or paradigm (Dosi, 1982)
146 (de Weck et al 2011)
147 (Lipsey et al, 2006; Winter, 2008; Funk, 2013a; Funk, 2013b)
148 (Utterback 1994; Adner and Levinthal 2001)
149 (Linton and Walsh, 2008)
150 (Hughes, 1984; Rosenberg, 1969; Dosi and Nelson, 2010)
151 (Green and Wind, 1979; Adner, 2002, 2004; Adner and Zemsky, 2005; de Figueiredo and Kyle,
2006)
152 (US DoE, 2010; Azevedo et al, 2009; Koh and Magee, 2006)
153 (Arthur, 2007)
154 (Basalla, 1988; Fleming, 2001; Fleming and Sorenson, 2001; Arthur, 2007)
155 (Arthur, 2007)
156 (Basalla, 1988; Fleming, 2001; Fleming and Sorenson, 2001; Arthur, 2007)
157 (Pavitt, 1984; Chandler)
158 (Schmookler, 1966)
159 See (Azevedo 2009) for LEDs, (Sheats et al 1996; Lee 2005) for OLEDs, and (Martinson, 2007;
Orton, 2009) for lasers.
160 See (Nemet 2006) for solar cells, (Shaw and Seidler, 2001; Dong et al, 2010) for organic
transistors, (NAS/NRC 1989) for load bearing materials, (Tarascon 2009) for batteries, (Bolund et
al, 2007) for flywheels, and (Simon 1995) for biological materials.
161 (Lipsey, Carlaw, and Bekar, 2005).
162 See for example (Chandler, 1994; Pratten, 1971; Gold, 1974; Freeman and Soete, 1997)
163 (Axelrod, Caze, and Wickham 1968; Mannan 2005).
164 (Keshner and Arya, 2004)
165 DisplaySearch, Flat Panel Display Market Outlook,
http://www.docstoc.com/docs/53390734/Flat-Panel-Display-Market-Outloo
166 (Keshner and Arya, 2004)
167 (Zhang,2011)
168 (Lipsey, Carlaw and Bekar, 2005; Winter, 2008)
169 (Lipsey, Carlaw, and Bekar, 2005)
170 Such a calculation assumes that production volumes are the same for all sizes of engines and that
there are no limits to increasing the scale of these engines. Since small engines probably have much
higher production volumes than larger engines the price data probably underestimates the benefits
from scaling while the extrapolations probably overestimate the benefits from scaling.
171 See (Hirsh, 1989) and in particular Figures 16 through 21. Increasing the voltage in transmission
systems dramatically reduced the energy losses in long-distance transmission and without low
energy losses, it would have been difficult to benefit from the geometrical scaling in generating
stations (Hirsh, 1989; Munson, 2005; Smil, 2010)
172 Data on capital cost per output is from (Hirsh, 1999). Edison’s Pearl Street Station Plant in 1880
was about 100 kw. Benefits from geometrical scaling can also be seen in the price per kilo-watt of
existing diesel generators. For example, the price of a large Cummins engine (2250 kw) is less than
20% that of smaller ones (e.g., 7 kw) on a per unit output basis. see the data on the following site:
http://www.generatorjoe.net/store.asp
173 See (UNCTD, 2006), Wikipedia’s entries on Airbus planes, and
www.airbus.com/.../media_object_file_2010-Aircraft-List-price.pdf
174 Like engines, there is greater demand for smaller aircraft and ships than smaller ones and thus
differences in demand are not driving these differences in price per output. See (UNCTD, 2006) and http://www.boeing.com/commercial/cmo/. The smaller benefits from scaling in aircraft than in oil tankers may be
because of the increasing cost of composites and other engineered materials in the largest aircraft,
which was also found to be a problem in wind turbines (See subsequent section).
175 (Glaeser and Kohlhase,. 2004)
176 See for example, (Kurzweil, 2005; ICKnowledge, 2005)
177 (Kurzweil, 2005; Teo, 2010: Humphries, 2010).
178 Other reasons include the falling cost of semiconductor lasers and camera chips for identifying
the bases with fluorescent dyes (Carlson, 2010) where these lasers and camera chips benefit from
both reductions in feature size and increase in wafer size
179 The notion that improvements in manufacturing equipment depend on the use of this equipment
to make parts for this equipment is also evident in computers; improvements in computers are
needed in order in for semiconductor manufacturing equipment to produce the ICs for the
computers.
180 (Freeman and Louca, 2001; Hounshell, 1984).
181 (Argote, L. and Epple, 1990; Huber, 1991; March, 1991).
182 One paper that does recognize this distinction is (Gold, 1981)
183 (Henderson and Clark, 1990)
184 (Daniel et al, 1999)
185 Although early analyses suggests that the cost of computing power only increased as the square
root of processing speed, more recent analyses and the replacement of large (e.g., mainframe)
with small (personal) computers suggest that this relationship is much less important than it was
originally thought (Ein-dor, 1985; 1997).
186 (Christensen, 1997; Christensen, Craig, and Hart, 2001; Christensen, Grossman, Hwang, 2008;
Christensen, Johnson and Horn, 2008)
187 For example, the Economist devoted at least five articles to Christensen and his ideas in 2010
and 2011
188 This is particularly true in Christensen’s most recent book with two co-authors (Dyer, Gregersen
and Christsensen, 2011)