Executive Summary
The inflation rate has grabbed headlines over the last several months by reaching heights not seen in the past four decades. The price of housing, vehicles, and gasoline have all risen dramatically, contributing to a higher consumer price index overall. One sector that has moderated rising prices more generally is the digital economy. Prices for goods and services that are either entirely digital, like mobile broadband and streaming video, or related hardware, like computers and smartphones, have not risen nearly as high as prices in the economy overall.
Despite their moderating influence on prices, the rapid development of new goods and services in the digital economy creates a lag within traditional measurements of inflation. New product models are not added to price indexes quickly enough to accurately account for improvements to quality or reductions in prices. This problem was exacerbated by the COVID-19 pandemic, which forced many people to rapidly adopt digital services like Zoom to work, study, and stay connected with friends and family. These errors in measurement for the digital economy tend to overstate the overall inflation numbers.
Correcting this mismeasurement of prices in the digital economy will require more frequent addition of new digital goods and services into price indexes. Adjusting how these prices are weighted in overall inflation figures would more accurately capture the digital economy’s role in moderating price increases. Finally, calculating these weights in parallel with the Economic Census data they are based on would reduce lag in updating inflation figures for the digital economy. The growing importance of digital goods and services in the overall economy makes it imperative that their prices are accurately captured in inflation statistics that drive U.S. economic policy.
Introduction
For three decades, the digital economy — defined as comprising information and communications technology (ICT) goods and services, online platforms, and e-commerce vendors — tended to have flat or declining prices, helping to keep inflation down. Although the behavior of prices in the digital economy has been affected by recent inflation, official and private price indexes for goods and services supplied by the digital economy show that inflation remains significantly lower in the digital economy than in the traditional economy. Thus, the digital economy has continued to play a moderating role in inflation.
Among the lasting deflationary influences on digital economy prices are effects of decreasing technology production costs, improving product quality, innovation to take advantage of this technology, and enhancements to competition as consumers gained access to e-commerce suppliers and suppliers on digital platforms. Cost reductions from the growth of global supply chains facilitated by digital technology have long contributed to lower prices both in the digital economy and beyond.
Research has found that the deflationary influence of the digital economy was even greater than indicated by the official indexes. One reason is that at times of rapid growth of the digital sector, the weights used to calculate inflation may not fully reflect the pace of digitization. For example, the pandemic greatly accelerated adoption of digital innovations such as e-commerce, so it’s reasonable to suspect that the price statistics are undercounting the impact of low digital inflation.
Feasible measurement improvements would likely help the main inflation indicators reflect the deflationary influence of digital economy prices more fully. These include increasing the representation of e-commerce prices in consumer price index (CPI) samples and updating the weights of the CPI annually. It is also important to reduce the lag in incorporating the results of the latest Economic Census in calculating the Producer Price Index (PPI), whose weights are still based on the 2012 Economic Census. These objectives may require legislation to facilitate incorporation of e-commerce prices in official indexes and collaboration in processing non-public data among the statistical agencies.
However, it should also be noted that recently global supply chains for semiconductors (and other products) have proven to be vulnerable to shocks from COVID, natural disasters, and rapid shifts in demand. These disruptions have led to problems in the supply of products as diverse as cars and appliances, contributing to the reemergence of high inflation.
Is the Digital Economy still helping to Moderate Inflation?
After almost four decades of relative quiet, inflation has soared. The 7.3% increase registered by the CPI-U in 2021 (as measured by comparing an average of the December 2021 and January 2022 indexes to an average of the December 2020 and January 2021) marks an abrupt end to a long era of moderate or low inflation. This period started with a large drop in the trend rate of inflation during the recessions of 1980-1982 and was followed by smaller declines in the recessions of 1990-1991 and 2008-2009 (Figure 1). The producer price index (PPI) for final demand also showed high inflation, with an increase of 9.7% in 2021. A third widely cited measurement of inflation, the personal consumption expenditures deflator from the national accounts, rose by 5.9% compared with 1.4% in 2020.
Note: Growth between year endpoints measured by as an average of adjacent December and January indexes)
Source: U.S. Bureau of Labor Statistics, All-items CPI-U
The recent resurgence in inflation has not affected all products equally. Unsurprisingly, attention has centered on the largest price increases, including those for energy (28% in 2021), used vehicles, (39%) and new vehicles (12%). More than half the 2021 acceleration in inflation can be attributed to energy and vehicles; if their prices had increased by 2% (the average growth rate of the CPI-U in 2016-2020), the rise in CPI in 2021 would have been 4.3% rather than 7.3%. Yet, in addition to asking where inflation was highest, it is also worth considering the segments of the economy that are continued to exert a moderating influence on inflation.
BLS Price Indexes for Digital Economy Products and Industries
Price declines were long common in the digital economy. 1 The recent re-emergence of inflation has changed price behavior throughout the economy, and the price indexes for digital products and industries included in the CPI, PPI or import price index (MPI) from the Bureau of Labor Statistics (BLS) have begun to show increases. Nevertheless, the digital economy continued to exert a moderating influence on inflation in 2021 as the price increases were relatively subdued and some prices continued to decline.
The digital CPI can be defined as including information technology hardware and services and wireless phone service, with each item weighted in proportion to its weight (or “relative importance”) in the CPI. The moderating influence on inflation of these products is evident, as the digital CPI rose just 0.6% in 2021 and fell in the first four months of 2022 (Figure 2). Falling prices of digital products also helped to keep inflation low in the pre-pandemic years. In the first year of the pandemic, however, the digital CPI increased 2.1% while overall inflation fell to 1.4%.
The data used to calculate the digital CPI can also be used to illustrate the potential for weights to have a significant impact. BLS introduces updated CPI weights in January of even-numbered years. The annualized rate of decline of the digital CPI in the four months beginning with January 2022 would have been 1.8% if the old weights had continued to be used, compared with 3.6% implied by the updated weights. Furthermore, the growth rate of the all-items CPI would have been 11.6% rather than 11.5%.
The CPI for information and information processing published by BLS differs from the digital CPI by including landline telephone service prices. It exhibits slightly more inflation, but its growth rate in 2021 is still low, at around 1% (Table 1). Furthermore, the products included in the digital CPI have low or moderate inflation rates in 2021. Prices declined in 2021 for telephone hardware and other information items, wireless telephone service, and software and computer accessories. Providers of internet and electronic information services had a moderate price increase of 2.6%. However, the CPI for computers and peripherals had a significant rise of 3.2% in 2021 driven by strong demand and a semiconductor shortage. This followed a 1.2% rise in 2020.
Source: Author’s calculations using price and relative importance data from BLS
Producer prices also showed almost no inflation in 2021 for major digital services, but digital equipment and components prices had significant increases at the producer level (Table 2). The PPI for computers and peripheral equipment rose slightly more that the CPI for computers. This increase was preceded by a 3.4% decline in 2020, however, so the two-year average rate of change of this PPI was just 0.3% per year. The broader PPI for computer and electronic product manufacturing, was up 3.4% in 2021, while the detailed PPI for semiconductor and other electronic component manufacturers was up 3.2% in 2021 with a two-year average rate of change of 0.2% per year. Slightly less inflation occurred in producers’ telecommunications equipment prices, which were up 2.8% after a history of being flat. The PPI uses hedonic techniques to quality-adjust prices of semiconductors and computers but not prices of telecommunications equipment.
The major digital services industries with flat PPIs in 2021 were wireless telecommunications carriers, software publishers, and suppliers of data processing, hosting, and related services. Furthermore, the PPI for internet access services fell 0.5%. Even the broader PPI for wired and wireless telecommunications showed well-subdued inflation, with prices up just 1.6%. However, three digital services had significant price increases. Video programming distribution via cable rose 3.7%, the price of advertising on digital platforms rose 5.6%, and the dollar margins received by electronic shopping industries rose by 4.2% in 2021 after rising by 7.9% in 2020. Retail dollar margins tend to increase with the price of goods sold, but with more volatility. With a constant cost of goods sold, a given percentage change in the price paid by shoppers would translate into a much greater percentage change in the margin for the retailer.
Import prices are also important to consider because they influence domestic inflation. Changes in import prices are at least partially passed through to the prices consumers pay. In the case of investment goods directly by a business, the import price is the price paid by the final user. The import price indexes (MPIs) are also of interest because they reflect developments in global supply chains, parts of which are under the control of multinational enterprises headquartered in the U.S. The import prices for computers and telecommunications equipment show little sign of inflation, as their MPIs rose less than 2% (Table 3). Nevertheless, the MPI for computer accessories, peripherals, and parts (which include semiconductors) rose by almost 6%, and the more detailed import index for semiconductors rose by 4.8%. The increases in prices of imported computer accessories and parts, which are in sharp contrast to their previous behavior, appear to reflect supply chain disruptions.
Online shopping
Advantages of convenience, selection, and lower prices have given e-commerce a steadily rising share of retail spending over the past 20 years. 2 The pandemic accelerated this trend, with e-commerce rising from 5% of non-gasoline retail spending in 2010 to 12% in 2019, and then jumping to almost 15% in 2020-2021 according to Census Bureau data. Adobe’s online marketing and web analytics unit found an even bigger jump in online spending of 42% in 2020. 3 The rapid growth of e-commerce is likely to cause under representation of e-commerce outlets in CPI outlet samples for products frequently bought online.
BLS does not publish CPIs for e-commerce prices, but information on e-commerce prices is available from the Adobe Digital Price Index (DPI). This index incorporates data on online transactions at 80 of the top 100 retailers. 4
The DPIs for detailed product categories reveal that the pandemic changed the behavior of online prices. Before the pandemic, consistent declines in online prices of electronics and computer components helped to give the overall DPI a downward trend. In 2020, however, the 12-month growth of the DPI turned positive in June and exceeded that of the CPI 5 in October and December (Figure 3). In 2021, the 12-month growth of the DPI accelerated to around 3% (and it went on to reach 3.6% in February 2022).
Nevertheless, an online inflation rate of 3% in 2021 remains significantly lower than overall inflation as measured by the all-items CPI. Furthermore, the top category for online spending, electronics, still had a falling DPI in 2021, and just four of eighteen product categories had significant higher inflation online than in the corresponding CPI (Table 4). The unweighted average of the price changes for the product categories with detailed DPIs is 2.7%, compared with 4.3% for the corresponding CPIs.
Most of the rise in the overall DPI in 2021 can be attributed to the 16.2% increase in online apparel prices, which have a relatively large weight in the DPI. Supply chain disruptions (including shipping bottlenecks and cost spikes) and a recovery of demand for apparel as pandemic restrictions eased contributed to the increase in online apparel prices. Also, part of the apparel price increase in 2021 was a retracing of an earlier price decline when demand for apparel was depressed by the pandemic: the two-year average growth rate of the DPI for apparel is just 6.2%. Finally, Goolsbee and Klenow (2018a) found a chain drift of 2.4% per year in earlier years in the DPI for apparel, suggesting that this source of index number distortion may have exaggerated the increase in the apparel DPI in 2021. 6
Figure 3: Adobe DPI compared with the All-Items CPI
Source:
Disinflationary Influence of the Digital Economy Prices in Earlier Years
Data on digital economy prices before the pandemic show the durability of the forces that have given the digital economy’s prices their current moderating role in inflation. Price indexes for computers and peripherals declined by double digits for many years starting in the early 1990s, and other ICT prices were also flat or declining after 1995. Around 2010, the rate of decline of prices of computers and peripherals and computer components slowed to the low single digits, perhaps reflecting a winding down of Moore’s Law (Flamm, 2021). 7 Inflation was still below zero in 2015-19 for most of the digital products in the CPI, PPI, and MPI.
The DPI, which begins in 2014, shows that e-commerce prices also declined in the years preceding the pandemic. Moreover, the detailed DPIs show lower inflation at e-commerce retailers than for the similar products at the outlets in the CPI. Goolsbee and Klenow (2018b) used DPI data to construct indexes of online prices for 65 “entry level items” of the CPI. The weighted average growth rate of the CPIs for those entry level items is -0.3% per year in 2014-2017, while the average growth rate of the corresponding DPIs is -1.6%. Methodological differences (the DPIs include new varieties as soon as they start to be purchased and use current spending shares as weights) could contribute to the 1.3 percentage point gap between the CPIs and DPIs but are unlikely to account for all of it.
In the case of apparel, information on Amazon’s prices is also available. An analysis by researchers from Amazon and academia suggests that the inflation rate for apparel from Amazon was well below the rate of change for the CPI for apparel from 2014 to 2017. To summarize, Amazon’s chief economist Patrick Bajari and academic co-authors (2021) constructed hedonic indexes of Amazon’s apparel prices where prices of items with non-matching characteristics are adjusted for quality differences so that they can be compared. The authors used artificial intelligence techniques to adjust for quality differences and tested two methods for calculating long run indexes. One method calculates month-to-month indexes and chains them together to calculate annual indexes and other longer run indexes. The other method constructs direct annual indexes that compare a month to the same month of previous year. Monthly chaining of the hedonic apparel indexes implied an average inflation rate of -5.3% per year from 2014 to 2017, while annual chaining implied an inflation rate of -0.77% per year. The annually chained index is unlikely to overstate deflation, yet its rate of change is three-quarters of a percentage point below the rate of change of the CPI for apparel, which is near zero. The Adobe DPI (which is matched models index with monthly chaining) is more consistent with the Amazon indexes, as its average rate of change in 2014-2017 is -1.3% per year.
The Unmeasured Deflationary Influence of the Digital Economy
Even though price indexes for digital products often fall, or at least fail to rise, the deflationary influence of the digital economy has not been fully captured in official measures of inflation. The first and perhaps biggest issue is that the weights and varieties used by the Bureau of Labor Statistics to calculate the consumer price index and the producer price index may significantly lag behind the rapid pandemic-driven digitization of the economy. Frequent introductions of new models and new products and rapid growth of e-commerce make inflation in the digital economy challenging to measure. BLS has procedures to adjust key ICT products for quality change, but for all the needed quality adjustments, new models and varieties must be brought into the elementary index promptly after they appear. That applies to new services as well as goods. Consider, for example, the dramatic surge in home and education usage of video conferencing programs such as Zoom, and the rapid shifts in consumer behavior around streaming content.
Importantly, e-commerce outlets could be significantly underrepresented in current CPI pricing samples. The BLS reports that it rotates the CPI outlet samples at 4-year intervals (BLS, 2018), which is unlikely to keep up with the rapid growth of online shopping during the pandemic. For consumer goods frequently purchased online today, that could mean that the measured inflation rate is overestimated.
More generally, outdated weights tend to dilute the influence on inflation estimates of fast-growing segments of the economy such as the digital sector. The CPI updates its higher-level weights at two-year intervals with a one-year lag from the end of the expenditure measurement period, so the average age of the weights over the cycle is three years. Under ordinary circumstances, that’s not much of a problem. But given the rapid pace of digitization during the pandemic, it could have a measurable impact. The effect of the weight update of January 2022 on the digital CPI above is an example.
The Producer Price Index poses a different issue. The scary headline numbers are generated by weighting individual industry and commodity inflation rates using data collected in the Economic Census every five years. The latest Economic Census was done in 2017, which already reflects a much different economy from today. However, the BEA has not yet published the detailed benchmark input-output tables for 2017.
The weights of the final demand PPI are currently based on the Economic Census of 2012 and input-output accounts developed from the 2012 Economic Census. 8 These PPI weights do not account for the growth in importance of the digital economy since 2012 or for the tendency of the falling relative prices of digital products to generate increased demand. As a result, the indicators used to adjust for changes in the economy since the benchmark year may underestimate the growth of fast-growing digital activities with many new entrants. The last time the benchmark year of the national accounts was updated, the annual growth rate of the digital economy over the preceding ten years was revised up from 5.6% to 7.8% per year (Jolliff and Nicholson, 2019).
In addition to underweighting of digital products in higher level aggregates and under-representation of e-commerce outlets in the CPI, researchers have identified other issues that could dilute influence of digital economy prices on key inflation indicators. These include: (1) under adjustment for quality improvements enabled by better technology, (2) outlet substitution excluding changes in the price that consumers are paying , and (3) the narrow measurement objectives of the CPI and GDP excluding some types of welfare gains from the growth of the digital economy .
These measurement problems are not new. To put them into perspective, Moulton (2018) updated the estimates in Lebow and Rudd (2003) and found that for all types of products and outlets in the CPI, the inflation rate mismeasurement due to under adjustment for quality change could amount to +0.37 percentage points, and inflation mismeasurement due to outlet substitution bias could amount to +0.08 percentage points. An analysis of the differences between personal consumption expenditure (PCE) weights and CPI weights also suggested that the inflation mismeasurement due to weighting problems in the CPI could amount to +0.1 percentage points.
However, the changes in outlet and product spending patterns in the period leading up to and including the pandemic may have been unusually rapid. In addition, BLS has continued to improve its quality adjustment techniques. It is therefore useful to consider these measurement issues separately.
Under adjustment for quality changes in ICT products
Advances in computing technology frequently enable quality improvements in ICT products. To account for these quality improvements, the sample of models and varieties used to calculate the price index must be regularly refreshed to include the new models and varieties as they appear and their prices must be adjusted for quality change.
Both steps are challenging. Official price indexes have therefore tended to under adjust for quality improvements in ICT products. For example, a quality-adjusted deflator for computers and peripherals estimated in a research setting falls 12% per year faster in 2004-14 than the official deflator, and a quality-adjusted deflator for communication equipment falls 7.6 percentage points faster than the official deflator (Byrne, Fernald and Reinsdorf, 2016). Furthermore, a quality-adjusted deflator for smartphones calculated for research purposes falls fast enough to reduce the growth rate of the personal consumption expenditure deflator by 0.05 percentage points in 2010-2017 (Aizcorbe, Byrne, and Sichel, 2019).
The speed with which new models and varieties enter CPI and PPI samples, and gaps in application of quality adjustment techniques may keep quality improvements in digital products from being fully captured. Nevertheless, the historical effect of overlooked quality improvements in digital products on the CPI growth rate appears to be modest, as indicated by Moulton’s (2018) estimates. Moreover, BLS has expanded its use of quality adjustment techniques, and in the case of smartphones in the CPI, the quality adjustment and sample refreshment practices in place since 2018 should fully capture the quality changes in smartphone models. The CPI and PPI also adjust for quality change in computers, peripherals, and internet access service; the CPI also quality adjusts other digital products. Furthermore, the weight of the digital products in the CPI is only around 6%.
Outlet substitution bias
E-commerce has fostered increased competition by giving consumers access to merchants outside their local area. Peer-to-peer platforms have also given consumers access to new kinds of service suppliers. The CPI does not track changes in the average price paid caused by changes in the outlets that consumers are buying from. The entry of lower-priced outlets can therefore lead to an overstatement of the change in cost of living, known as outlet substitution bias. Online prices tend to be slightly lower than offline prices (Cavallo, 2017, reports that prices are 5% lower on Amazon than offline), so consumers have been able to pay lower prices by shopping online. Estimating the impact on the cost of living of consumers’ savings from paying lower prices online would require correctly measuring the shift to e-commerce and the effect of this shift on prices that consumers are paying.
Peer-to-peer platforms for services such as ridesharing and short-term rentals have also enabled consumers to pay lower prices. Research on substitution to ridesharing has found a substantial decrease in the average cost of trips between fixed pairs of locations. Comparing a pooled index for taxis and ridesharing with an index for taxis alone showed that a decline in the average price paid caused by substitution of ridesharing for taxis subtracted 0.6 percentage points from the growth rate of the pooled index in New York City from Q3 of 2015 to Q4 of 2017 (Aizcorbe and Chen, 2022). Ridesharing’s lower inflation rate subtracted an additional 0.4 percentage points from the pooled index’s growth rate. A downward correction of 1 percentage point to the growth rate of the CPIs for both intra-city transportation and lodging away from home would reduce the growth rate of the all-items CPI by a small amount.
Other impacts on consumer welfare
Broader aspects of consumer welfare could be included in a cost-of-living index and are relevant for understanding the economic effects of the growth of the digital economy.
For example, the pre-pandemic shift to e-commerce happened over the same period as a sustained decline in household shopping time (including travel) as measured by the American Time Use Survey. According to Mandel (2020), per capita unpaid hours devoted to non-gasoline non-grocery shopping fell by 27% between 2007 and 2018. This decline reduced unpaid shopping time by 10.5 billion hours in 2018.
Once the pandemic started, this trend accelerated. Time spent shopping for all consumer goods dropped by roughly 25% in 2020 compared to 2019 (for both years, May to December). That includes time spent online, according to the latest release of the BLS American Time Use Survey. When the 2021 ATUS data is released, it will be interesting to see the long-term changes in consumer behavior.
As the shopping example shows, estimates of welfare gains from changes in consumer behavior can be sizeable. One strand of the research on the growth of consumer welfare from the rise of the digital economy finds that including the consumer surplus from free digital platforms such as Facebook significantly raises the growth rate of an extended version of GDP known as GDP-B (Brynjolfsson et al., 2019). Another study finds adjustments for the value to consumers of the expanded selection and convenience offered by e-commerce would significantly reduce the growth rate of an extended version of the PCE deflator (Dolfen et al., 2019). Finally, the measures of output growth and inflation in national accounts do not reflect the rapid growth in consumers’ utilization of digital communication and entertainment services (internet access, mobile telephony, cable, and streaming of videos and music). Adjusting the prices paid by consumers for their utilization of these services as measured by data usage and talk time reduces the rate of change of the PCE deflator after 2008 by almost a half percentage point (Byrne and Corrado, 2020). 9
The Semiconductor Shortage
Prices of digital goods and services are the main way the digital economy influences inflation. However, a non-price effect has also recently become important: the shortage of semiconductors. Over the past 20 years, new applications of digital technology based on semiconductors have enabled quality improvements in a wide range of products, putting downward pressure on prices. Most semiconductors are, however, produced in fragmented global supply chains that are vulnerable to shocks and that may lack flexibility to respond to sudden changes in demand. Disruptions to supply chains have led to shortages of the semiconductor chips needed to produce many types of products, including medical devices and broadband equipment.
10
Vakil, Bindiya, and Tom Linton. “Why We’re in the Midst of a Global Semiconductor Shortage.” Harvard Business Review, February 26, 2021. https://hbr.org/2021/02/why-were-in-the-midst-of-a-global-semiconductor-shortage.
Semiconductors are not included in the widely used gauges of inflation, and they account for only a small part of the production cost of most products that contain them. Nevertheless, the unavailability of an intermediate input for which there is no substitute can have a disproportionate effect on final demand prices. The effect of rationing on a buyer’s price index can be modeled using the hypothetical price that would cause the quantity demanded to equal the rationed quantity. In the case of an irreplaceable input to production, this virtual price would be quite high.
The most important effect of the semiconductor shortage on the CPI has come from the impact on markets for new and used vehicles when the shortage forced auto makers to curtail production. The price increases for new and used vehicles contributed about 1.5 percentage points to the change in the all-items CPI in 2021. Part of this contribution is attributable to the semiconductor shortage (though factors such as fluctuating demand for rental cars also contributed).
Policy Implications
The flat or falling prices of the digital economy long exerted a deflationary influence on the US economy. These prices continue to play a moderating role in inflation after the recent reemergence of inflation.
Given the value of technology in moderating inflation, policymakers should seek opportunities to promote uses of digital technology to enhance growth and promote price stability. At the same time, the contribution of the semiconductor shortage to inflation shows the need for supply chain improvements both to reduce current inflationary pressures and to reduce risks of future shocks. During the era of globalization, growth of international supply chains that located each step of production where costs were lowest had deflationary effects, but also increased vulnerability to inflation-generating supply disruptions. These risks have recently been exacerbated by COVID and other changes in the international environment.
The focus on inflation in the current economic debate also highlights the importance of accurate measurement of price change. Wider application of timely sample refreshment procedures currently used for smartphone models in the CPI could help to capture quality change more fully in the CPI and PPI. 11 In particular, incorporating scanner data and data on e-commerce transactions in the CPI would help to bring new models and products into the elementary indexes more quickly (National Academies of Sciences, 2022). The e-commerce transactions data would also help to improve the representation of e-commerce prices in the CPI. The practical obstacles posed by the size and confidentiality concerns of the e-commerce transactions-level data sets must first be overcome, however. Having the e-commerce merchant calculate indexes or index components with procedures specified by BLS would solve both problems. However, doing these calculations involves more effort than responding to a standard statistical agency survey, so legislation may be required to delegate this responsibility to merchants.
Another measurement challenge where progress should be possible is correctly weighting digital products in top-level indexes of consumer, producer, and national accounts prices. Annual weight updates assisted by information on personal consumption expenditures from the national accounts would improve the timeliness of the CPI weights.
Shortening the six-year lag in benchmarking the national accounts to the new Economic Census would improve the PPI’s measurement of current inflation and the national accounts’ measurement of the digital economy. The lag could be shortened if the work on the national accounts benchmark estimates could proceed in parallel with the processing of the responses to the Economic Census. However, this would require either unifying the organization of U.S. statistical system or enhancing the ability of the statistical agencies to share non-public data with each other.
Conclusion
The promise of the digital economy is a constantly advancing suite of goods and services provided by information and communication technologies. Adjusting inflation figures for the quality of smartphones, computers, streaming media, and e-commerce is vitally important to measuring their contribution to the economy overall. Collecting price data about new goods and services must keep pace with their introduction. This is especially important when a pandemic accelerates the growth of the digital economy. Having up to date prices for information and communications technology will give a more detailed picture of overall inflation.
Four decades of relatively low inflation have made the current inflation figures particularly troubling. Over this same time, advances in the digital economy have greatly increased the speed of communication and access to information, saving Americans substantial amounts of time and money. Given these continuous improvements, recent inflation figures may be overstated. More accurate price reporting would help economic policymakers and consumers take action to counteract inflation’s effects on government and household budgets.
Report–American Science and Technology Leadership Under Threat: Restrictive Antitrust Legislation and Growing Global Competition
Executive Summary
- The package of antitrust legislation recently introduced in Congress would limit tech companies’ ability to integrate new products, promote new features, and compete in new market segments. Antitrust regulations that reduce the size and limit the scope of tech firms weaken their incentives to make the large-scale, long-run investments in science and technology, vital for national security and economic prosperity.
- At a time when the United States critically depends on a handful of firms to pursue large scale research projects, such proposals would play into the hands of foreign rivals. U.S. leadership in four of the five emerging technologies identified by the Biden Administration as a national priority, would be adversely impacted by proposals to restrict digital platforms.
- The research laboratories of leading corporations are an important source of scientific advances and technical breakthroughs. They are vital for applying these advances in the development of commercially viable products and processes.
- Some of the leading technology firms continue to invest in scientific research even though others are withdrawing. These firms must be nurtured and encouraged to continue to develop such general-purpose technologies as quantum computing and artificial intelligence.
- Companies are under pressure to realize the financial returns of their general-purpose technologies allowing them to operate in a diverse set of markets. These firms need the incentives of secure returns on investments to continue their own scientific research.
- Antitrust measures can come with unintended consequences that will hurt American leadership in science and technology diminishing our national security and economic prosperity.
1 Background
The U.S. technology sector is facing barriers to its ability to advance its position in an increasingly competitive world. The package of antitrust legislation introduced in Congress (summer 2021) may adversely affect the tech sector by limiting the ability of platforms to design new products, integrate existing ones and operate in downstream segments.
In this report we highlight the potential impact of these limitations on American science and technology leadership. We examine the role that big firms play in advancing U.S. technology, the foreign competition they increasingly face, and the fragile nature of the U.S. innovation ecosystem.
2 Introduction
America’s leading-edge innovation and science-based breakthroughs drive U.S. economic prosperity and even support the national security of the nation. Our universities and startups are envied around the world and attract the world’s leading researchers. But the vital underpinning framework for the American innovation ecosystem includes leading corporations. Scientists from corporate labs collaborate with university researchers and startups to produce crucial research that creates new general-purpose technologies. These R&D intensive corporations are working to develop, scale up, and commercially deploy the many applications of general-purpose technologies.
Companies are under pressure to realize the financial returns of their general-purpose technologies that can be applied to a diverse set of markets. These firms need the incentives of secure returns on their investments to continue their own scientific research. Antitrust measures can come with unintended consequences that hurt American leadership in science and technology diminishing our national security and economic prosperity.
The United States is losing its historic lead in innovation. China is quickly catching up in terms of both R&D investments and R&D intensity, its domestic expenditure on R&D as a percentage of gross domestic product. Chinese firms have narrowed the gap and even taken the lead in some segments that are critical for global technological leadership. Policy makers need to be cautious not to weaken the incentives of big American corporations to invest in research.
U.S. corporate research is increasingly concentrated in only a few industrial sectors (such as information and communication technologies and pharmaceuticals) and funded by only a handful of firms that still have the big scale and scope to make large, risky investments in scientific research. Legislation that prevents these firms from developing new products or entering new markets will reduce their investments in research and thus further undermine U.S. leadership.
The erosion in American leadership is not uniform — it is smaller in segments where American companies have made significant investments in scientific research, and greater in those where such investments are limited. This report shows that U.S. firms maintain global leadership in tech segments based on leading-edge science such as artificial intelligence and quantum computing. However, foreign competitors have caught up in tech segments based on more mature science (such as 5G telecommunications). Therefore, corporate research is vital for preventing further erosion.
Science and technology are intertwined in complex ways, so the line between these activities may be difficult to draw at times. Nevertheless, doing so is important to truly understand the U.S. innovation ecosystem. Science is a systematic enterprise directed towards obtaining new knowledge about the world. Scientific knowledge is concerned with general laws, obtained through basic and applied research (the “R” or upstream part of R&D), and typically disseminated broadly in scientific literature (e.g., publications and conference proceedings). Technology is the application of scientific knowledge for practical purposes. Technical knowledge is concerned with how and why specific artifacts work, obtained through development (the “D” or downstream part of R&D), and typically protected as intellectual property (e.g., patents). It is important to understand these distinctions before delving into the details of the innovation ecosystem.
3 The United States Is Losing Ground in Innovation
The United States is losing ground in innovation despite having been at the forefront of science and technology for many years. Its universities and corporations have garnered Nobel Prizes in physics, chemistry, physiology, and medicine. American corporations, undisputed global technology leaders such as Google, Intel, IBM, and Microsoft, have created entire new industries. Today, they occupy such science-based industries as artificial intelligence and quantum computing. It might appear that American science and technology reign supreme. Yet, there are causes for concern.
This can be seen when looking at key innovation inputs (e.g., R&D expenditure, human capital, venture capital, and scientific research) as well as key innovation outputs (e.g., patents, products, and the resulting foreign trade balances). On the input side, other countries have stepped up their investments in science and technology:
- China almost quadrupled its R&D intensity from 0.57% in 1995 to 2.2% in 2019 [as can be seen on Slide 7, Figure 1]. 1
- The United States experienced a modest increase over the same period, from 2.4% in 1995 to 3.1% in 2019, and as a benchmark, the 27 countries of the European Union (EU) increased their R&D intensity from 1.57% in 1995 to 2.1% in 2019.
- While China’s R&D intensity still lags behind South Korea, Japan, and the United States, the country has quickly become the world’s second-largest R&D spender, with its total R&D expenditure reaching 84% of that of the United States in 2019 in purchasing power terms, up from 26% in 2005 [Slide 8, Figure 2]. 2
- China’s total factor productivity at constant national prices was 89.6% larger in 2019 compared to 1980. 3
- In the United States, total factor productivity was only 29.4% larger in 2019 compared to 1980. 4
Another crucial factor that influences the erosion of U.S. leadership in innovation is human capital. Between 2012 and 2021, the Chinese government doubled its investment in higher education. 5
The number of PhD students entering STEM fields at Chinese universities increased nearly 40% between 2016 and 2019. By 2025, China is projected to reach nearly double the number of annual STEM PhD graduates compared to the United States. Combined with the fact that a large share of Chinese PhDs graduate from high-quality universities, the growing quantity of doctoral graduates is a leading indicator of the country’s increasing competitiveness in science and technology.
Venture Capital (VC) also plays an important role in innovation by supporting startups that translate scientific research into products that can be commercialized. Global VC investments are growing rapidly but the United States is attracting a declining share of them. Data from the National Science Foundation (NSF) indicate that in 2005 U.S. startups garnered 80% of global VC investments, but only 44% in 2018. 6 Conversely, China increased its share from negligible levels in 2005 to about 36% in 2018 [Slide 10, Figure 3].
The United States remains the global scientific leader for scientific research, an essential input into technology development, though China has narrowed the gap significantly.
The recent trends in highly cited publications points to China catching up. The “index of highly cited publications” indicates the relative impact of a country’s scientific research and is a standardized score equal to the ratio between, (1) the country’s share of publications among the world’s top 1% of cited articles and (2) the country’s share of all publications.
- The U.S. index increased from 1.77 in 2000 to 1.88 in 2016, 7
- Over the same period, the EU index grew from 1.0 to 1.3,
- China’s index more than doubled, from 0.4 to 1.1.
The increase in China’s index highlights the rising impact of Chinese science [Slide 11, Figure 4]. The United States’ leadership position in science is also reflected in the use of science for innovation. For instance, the American share in ICT patents granted by the U.S. Patent and Trademark Office (USPTO) declined from 52% in the 1980s to 46% in the 2010s. However, when considering science-based patents, the American share remained stable at around 60%. 8
On the output side, China overtook the United States in Patent Cooperation Treaty (PCT) patent applications in 2019, when it filed over 118,000 applications compared to 115,000 from the United States [Slide 12, Figure 5]. 9 PCT applications allow applicants to seek patent protection for an invention simultaneously in a large number of countries. PCT applications, because of their added cost and complexity, tend to indicate that the underlying technology has broad potential application around the globe.
Innovation is ultimately embodied in new products and services. Foreign competitors, with greater government support and access to cheaper capital, have been able to catch up to and even surpass American firms by developing new products and services. In the information and communication technologies (ICT) sector, the United States had 55% of the 500 top-ranked supercomputers in 2010, but only had 23% in 2019 [Slide 13, Figure 6]. 10 Over the same period, China’s share of supercomputers increased from 8% to 44%. In semiconductors, American firms currently account for only 12% of global manufacturing, while Taiwanese, South Korean, Chinese, and Japanese firms account for a combined 73%. 11
The ability to retain domestic manufacturing capability is crucial for economic prosperity and national security. Data from NSF show that the United States has a steadily increasing trade deficit in high R&D intensive products, consistent with an increase in global high-tech competition [Slide 14, Figure 7] 12 . This deficit grew from $130 billion in 2005 to over $300 billion in 2018 (in current dollars). 13 While the United States has grown increasingly dependent on foreign imports of high-tech products, the same cannot be said of China. During 2005-2018, China has run a trade surplus in high R&D intensive products in all years except one.
The indicators of innovation outcomes show that the American lead is diminishing, whether measured by patents or the growing trade deficit in high-tech products. In short, the United States has lost ground in innovation relative to other countries, especially China. Its investments in R&D and STEM PhDs lag those of rivals. The United States still has retained its competitive advantage in its scientific leadership. Being on the scientific frontier allows American firms to invent new general-purpose technologies and pioneer novel commercial applications. However, as a field matures these advantages shrink.
4 American Leadership Remains Strong Where American Firms Invest in Scientific Research
The erosion in American leadership is not uniform. To understand where the United States remains the global innovation leader and where other countries are catching up, we have analyzed six technology segments identified as a national strategic priority by the Biden Administration: 14 quantum computing, artificial intelligence, 5G, semiconductors, cybersecurity, and autonomous vehicles.
American firms continue to lead in research-intensive emerging technology segments, including quantum computing and AI. However, the competition is closing in rapidly, particularly in AI. Helped by large-scale government support, Chinese firms and universities are making rapid progress in science and technology. For example, Baidu, Huawei, and Alibaba have increased their publication rate in AI since 2013 [Slide 22, Figure 11]. Moreover, Chinese organizations won all first and second places in all tasks of the 2021 AI City Challenge. 15 Additional details on science, technology, and innovation in quantum computing and AI are available in the Supplementary Materials section of this brief.
American firms are still responsible for significant scientific and technological breakthroughs in AI. Google, Microsoft, Facebook, and IBM are leading publishers of scientific papers in leading global AI conferences [Slide 21, Figure 10]. Foreign firms are lagging (with Chinese firms coming in a distant second place). Microsoft, Google, IBM, Samsung, Intel, Nokia, and NTT publish the most AI research published in all outlets [Slide 22, Figure 11]. On the technology side, U.S. firms account for 88% of (quality adjusted) AI patents granted by the USPTO to the top 50 organizations during 2006-2020. Moreover, American firms’ lead in AI can also be seen in marketplace innovations [Slide 24]. The United States is home to 103 of the 183 active AI “unicorns” (i.e., startups with a valuation above $1 billion).
American firms are also responsible for significant scientific and technological breakthroughs in quantum computing. Between 2013 and 2020, Google, Microsoft, and IBM together produced more quantum computing publications than MIT [Slide 18, Figure 8]. 16 Over the same period, IBM, Northrop Grumman, and Microsoft each received more (quality adjusted) quantum computing patents than MIT [Slide 19, Figure 9]. 17 Not surprisingly, firms that actively publish and patent in quantum computing are also pioneering quantum computing products and services [Slide 20]. For example, Google’s Quantum Computing Service allows users to run quantum programs and experiments remotely on Google’s latest production quantum processors, such as Sycamore, which has up to 54 superconducting qubits. 18
Foreign firms have taken the lead in several research-intensive mature technology segments, including 5G telecommunications and semiconductors. China’s Huawei, Finland’s Nokia, and Sweden’s Ericsson have led in 5G publications during 2013-2016 and 2017-2020 [Slide 27, Figure 13]. Moreover, Chinese firms Huawei, ZTE, and CATT increased their patenting activity in 5G between 2015-2020 [Slide 28, Figure 14]. As a result of its leadership position in 5G science and technology, Huawei has contributed 23% of the 5G standard patents, a significantly larger share than Qualcomm’s 6%. 19 Added to this, foreign firms that actively publish and patent in 5G are also leading in the product market. In 2019, Huawei earned 28% of global telecommunication equipment revenues, followed by Nokia (16%), Ericsson (14%), ZTE (9%), and Cisco Systems (7%) [Slide 30, Figure 16]. 20
The United States has lost semiconductor manufacturing: in 2020, it had only 12% of global chip manufacturing, compared to 22% for Taiwan, 21% for South Korea, 15% each for Japan and China, and 8% for Europe [Slide 35]. 21 By 2030, America’s share is predicted to fall to 10%, while China’s share is predicted to rise to 24%, with 77% of global fab capacity projected to be located in Asia.
Conversely, American firms still have a lead in semiconductor science (in contrast to manufacturing). Between 2011 and 2019, IBM, AT&T, Intel, Microsoft, Micron Technology, Google, and Global Foundries together produced slightly more publications than Samsung, Nokia, Toshiba, Taiwan Semiconductor, and LG [Slide 31, Figure 17]. While U.S. firms lead in semiconductor design, foreign firms are strong in semiconductor process technologies. For example, among the top 50 organizations that were granted semiconductor patents by the USPTO during 2006-2020, 42% were from Japan, 28% from the United States, 12% from Taiwan, and 6% from South Korea [Slide 32, Figure 18]. In general, semiconductor patents granted to American organizations have higher quality, but the foreign competition is strong. Both TSMC and Samsung surpassed IBM in the number of quality-adjusted semiconductor process patents during 2015-2020 [Slide 33, Figure 19]. On a positive note, U.S. firms that invest in scientific research in semiconductors continue to innovate in the marketplace. For example, in 2020, Intel led the world in semiconductor revenue ($74 billion), followed by Samsung ($60 billion) and TSMC ($45 billion) [Slide 34, Figure 20]. 22
American leadership is being challenged in other technology segments, including cybersecurity and autonomous vehicles. The United States narrowly leads in cybersecurity research, with foreign universities and firms having gained ground. IBM (444 papers), Microsoft (416 papers), and AT&T (380 papers) are in the top 10 of organizations based on the number of papers presented in the top 10 cybersecurity conferences. At the same time, Tsinghua University in China (621 papers) is second only to the University of California (1,367 papers). Moreover, Huawei (186 papers) has surpassed Intel (148 papers) and Google (144 papers), with Ericsson (116 papers) following closely [Slide 37, Figure 22]. On the technology side, U.S. firms received a majority of cybersecurity patents between 2005-2020. For example, IBM, Intel, and Microsoft received a combined 21% of all cybersecurity patents granted by the USPTO to the top 50 organizations [Slide 38, Figure 23]. Yet, important developments in cybersecurity may be kept as trade secrets, rather than disseminated through publication and protected through patenting.
Foreign organizations have also narrowed the gap with the U.S. in autonomous vehicle research (AV). For example, the largest publishers in the top 10 autonomous vehicle conferences are five U.S. universities (University of California, Carnegie Mellon University, MIT, Stanford University and University of Michigan), one U.S. firm (Microsoft), and four foreign universities (France’s CNRS, Switzerland’s ETH Zurich, and Japan’s Waseda University and Nagoya University). The firms included in the top 50 of conference publishers are Microsoft (#8), IBM (#14), Google (#34), AT&T (#43), and Intel (#44). Outside the top 50 organizations, Chinese firms Tencent (#59), Huawei (#61), and Baidu (#65) publish as much as Facebook (#51), NVIDIA (#56), and Amazon (#63) [Slide 39, Figure 24]. In terms of AV technology, automotive firms lead in patenting. Ford, GM, and Toyota received a combined 2,013 patents, representing 38% of all patents granted by the USPTO in 2005-2020 to organizations in the top 50. At the same time, Google, Microsoft, Intel, Qualcomm, and IBM each produced more AV patents during 2005-2020 than such leading foreign automotive firms as Honda and Hyundai [Slide 40, Figure 25].
In summary, multiple indicators demonstrate sustained U.S. leadership in innovation, including scientific publications, patents, product introductions, and market share. Yet they all point in the same direction — America is strong in industry segments where U.S. companies invest in leading-edge scientific research. In addition, market leading firms also actively participate in scientific research published in scientific journals and technology development measured by patents.
5 Corporate Research Is Vital for America’s Technological Leadership
The research laboratories of leading corporations are a crucial source of scientific advances and technical breakthroughs. However, U.S. investments in R&D, both private and public, are increasingly concentrated in a few industries, and funded by a small number of private firms [Slides 46-47, Figures 27,28]. As a result, American technology output is increasingly dependent on a small number of companies primarily in the information technology sector. Even there, America’s share of patents is shrinking, with the exception of science-based ones [Slide 50, Figure 31].
In both the United States and China, ICT R&D expenditures represent the largest share of manufacturing industries’ R&D investments. R&D in the ICT sector accounts for nearly a third of all business R&D in manufacturing in the United States. In 2018, business invested nearly $227 billion in manufacturing R&D, with ICT accounting for over $112 billion. In China, ICT accounts for about 22% of overall R&D in manufacturing but has grown at a much faster pace. Between 2011 and 2018, Chinese investment in ICT R&D has grown two and half times faster than in the United States, at about 13% per year, compared to 5.4% per year in the United States. 23
Survey series available at https://www.nsf.gov/statistics/srvyberd/. National Bureau of Statistics of China, Basic Statistics on R&D Activities of Industrial Enterprises Above Designated Size by Registration Status, from National Data series available at https://data.stats.gov.cn/english/easyquery.htm?cn=C01.
Research laboratories in corporations are vital for applying these advances in the development of commercially viable products and processes. The American innovation ecosystem is heavily reliant on private investment. National Science Foundation data shows that the ratio of federal funding for R&D to business funding for R&D was 2:1 in the 1950s. Since then, federal support for R&D has steadily declined. In 2018, the ratio was barely 1:4. 24 [Slide 45, Figure 26].
Although universities are responsible for performing the bulk of scientific research, the research performed by businesses is vital. Businesses accounted for over 45% of all basic and applied research performed in the economy in 2018. Their significance is even greater, as corporate research has been the basis of many technical advances that have profoundly changed our lives. For example, the processors that run today’s computers consist of integrated circuits pioneered by Jack Kilby at Texas Instruments, and integrated circuits are collections of transistors, which were invented by Bell Labs scientists. The laser beams that encode Zoom calls and the optical fibers that carry the information around the world were researched by TRG, GE, IBM, Bell Labs, and Corning. The graphical user interface that frames the Zoom window was based on research conducted at Xerox PARC [see Slide 51 for more examples].
6 General Purpose Technologies Go Hand-in-Hand with Big Size
There is a close relationship between the incentives to invest in research and the scale and scope of the firm. Without the leadership of firms with substantial scale and scope, the full potential of general-purpose technologies may not be realized. Investments by private corporations are ultimately guided by financial returns, as they should be. This is true of investments in research as well. But research is a risky enterprise. The returns are uncertain, and can take many years to materialize, if at all, and there is the ever-present danger that the benefits may not add value to the firm’s current products. Conversely, successful research investments will frequently result in an increase in scale and scope, as new product categories are introduced and entirely new markets open up. Scale and scope can, under enlightened and far-sighted management, deliver substantial investments in research. Down the line, these investments can yield growth and the creation of new markets and create greater scale and scope. 25
General-purpose technologies have broad applicability. A company that successfully creates a general-purpose technology will naturally try to apply it in different markets, and also try to integrate existing products based on a shared general-purpose use. Significant innovations based on general-purpose technologies are often complex, require advances in many individual components, and draw upon multiple scientific and technical domains. Naturally, this entails extensive coordination. Investments in such innovations take time to come to fruition, often encounter roadblocks, and carry significant technical and commercial risk. Large, diversified firms are better able to manage such risk. When successful, these investments bring substantial financial rewards to the innovator in the form of large market share and the ability to diversify and grow previously unrelated products.
Information and communication technologies have been characterized as general-purpose technologies (GPTs) due their broad applications within and across industries. For instance, artificial intelligence (including machine learning) is useful in applications as diverse as choosing a movie, ad displays, search queries, fraudulent credit card transactions, voice assistants, power grid data, self-driving cars, 3-dimensional protein structure, and discovering new cures for diseases. The ability to spread the cost of R&D across many different applications makes investment in GPT attractive for both private firms and society at large. Broadly speaking, information technology is a GPT. Reflecting its general-purpose nature, the top AI patenting companies include IBM, Google, Microsoft, Facebook, and Amazon, and also software producers such as Oracle, hardware manufacturers like Cisco, as well as financial services firms including Capital One and Bank of America.
General-purpose technologies combine knowledge from distinct scientific fields. Industrial labs are organized around technical problems, while university labs are organized by the rigid boundaries of scientific areas and academic departments. Arguably, the most important general-purpose technology of the twentieth century — the transistor — was invented by the largest and most renowned American corporate lab. Bell Lab’s advantage over universities and startups laid in its ability to attract and organize world-leading scientists around a common mission. AT&T granted Bell Labs with the patience and resources needed to accomplish major scientific breakthroughs in multiple research fields. It is hard to imagine the American innovation ecosystem being able to invent the transistor in the absence of its most successful corporate lab.
Artificial intelligence (AI) is a good example of the comparative advantage of the corporate lab in general-purpose technologies. Protein structure prediction has been a major challenge in biology for over 50 years. The social value of this research is high because the ability to accurately determine the 3D shape of proteins from their amino-acid sequence could greatly facilitate drug discovery. However, its private value is uncertain and lies many years in the future. DeepMind, a leading AI firm owned by Alphabet, developed a program called AlphaFold that outperformed 100 other teams in a protein structure prediction challenge in November 2020. Observers argued that the program could revolutionize biology in the long run. Just a month later DeepMind reported $649 million in losses for 2019, slightly larger than its losses in 2018. Despite this, Alphabet has pledged to keep funding DeepMind even though its accumulated losses were well in excess of $2 billion through 2020. 26 Unlike small firms, such large corporations as Alphabet and Microsoft can use AI research on a variety of internal projects, ranging from managing the power grid of data centers to improving the AI of voice assistant technology. This broad scope of applications is linked to the general-purpose nature of AI. Further, machine learning also requires substantial and expensive computing resources that are only available to large, well-established firms. Yet, the investment could pay off. Should DeepMind succeed in its quest to develop a “general” artificial intelligence, many new commercial applications could open up, increasing the scale and scope of its parent company, Alphabet.
Firms with scale and scope that made substantial investments in scientific research were key players in the rise of American leadership in science and technology. Over the last four decades, the American innovation ecosystem has become more diverse, with universities shouldering more of the responsibility for research, and startups becoming important in testing the commercial applications of scientific and technical advances. The great successes of this division of labor in innovation should not blind us to the threats to America’s leadership in science and technology. In particular, corporate labs remain vital to the American innovation ecosystem.
Mission-oriented research driven by commercial application of substantial scale and multidisciplinary scope is crucial when it comes to advancing general-purpose technologies. Firms invest in research to develop market-leading innovations, which if successful, lead to the creation of new products and market segments. To invest in large scale scientific research, and to create market-leading innovations, tech firms must be permitted to pursue large commercial scale and a wide range of product applications.
7 Established Firms Are Withdrawing from Scientific Research. The Few That Continue to Invest Should Be Nurtured.
The decline in corporate research is reflected in the downsizing or even closure of many corporate labs. In the early and mid-twentieth century, the DuPont Central Research & Development Organization was run on par with top academic chemistry departments. However, in the 1990s, DuPont’s attitude toward research changed as the company started focusing on the business potential of projects over basic research. As a result, the number of journal articles authored by DuPont scientists fell from 749 to 245 between 1994 and 2015, while the number of patents the company filed with the USPTO increased from around 1,600 in 1994 to close to 3,500 in 2012. This change reflected a shift to downstream development activities. Following pressure from activist investors, on January 4, 2016, DuPont’s Central Research & Development Organization ceased to operate as an independent research unit and was merged with DuPont’s engineering division.
The decline of the corporate lab was striking especially when considering that America’s public corporations increased in size substantially during this period. For example, net sales at GE grew from around $25 billion in 1980 to around $100 billion in 1998. However, employment of doctorate holders at GE’s corporate research laboratory dropped from 1,649 in 1979 to only 475 in 1998. Similarly, IBM’s net sales grew from $26 billion in 1980 to $82 billion in 1998. Yet, the number of doctorate holders working for IBM dropped from 1,300 in 1979 to 1,200 in 1998. 27 28
Over the past three decades the composition of business R&D changed to include less “R” and more “D” [Slides 52-55, Figures 32-34]. The share of research, both basic and applied, in total business R&D expenditures in the United States fell from about 30% in 1985 to below 20% in 2015. The changing composition of corporate R&D can also be seen by looking at R&D outputs: for publicly traded U.S. corporations with at least $10 million of R&D stock, scientific publications (a measure of upstream R&D) approximately halved, from around 20 per company per year in 1980 to around 10 in 2015. 29 By contrast, patents (a measure of downstream R&D) increased from around 10 patents per company per year to over 70 during the same period. The share of corporations that published more than 10 articles per year dropped from 55.2% in 1980 (111 out of 201 firms) to 29.8% in 2015 (214 out of 717 firms). The average number of scientific publications per $1 million of R&D spending also declined from 0.46 articles between 1980 and 1985 to 0.40 articles between 2010 and 2015. The decline was evident even among the fastest growing firms. For instance, U.S. public corporations whose sales grew by 100% or more between 1980 and 1990 experienced on average a drop of 20.6 scientific articles per year over that decade. Similarly, corporations that doubled in sales between 1990 and 2000 published 12.0 fewer articles over that decade, and publications dropped by 13.3 for the fastest growing U.S. public firms between 2000 and 2010. 30 In terms of research quality, the decline was most pronounced for highly cited papers and for papers published in high-impact scientific journals.
Only a handful of large firms continued to invest in internal scientific research [Slide 54, Figure 34]. Corporate research drives breakthrough innovation which cannot be done universities and startups. These firms must be nurtured and encouraged to continue to develop such general-purpose technologies as quantum computing and artificial intelligence. American leadership in IT and digital innovation rests on the willingness and ability of a few ICT firms to remain engaged in scientific research.
8 Conclusion
The package of antitrust legislation recently introduced in Congress would limit tech companies’ ability to integrate new products, promote new features, and even compete in new market segments. Antitrust regulations that reduce the size and limit the scope of tech firms weaken their incentives to make the large-scale, long-run investments in science and technology, vital for national security and economic prosperity.
At a time when the United States critically depends on a handful of firms to pursue large scale research projects, and given the dangerously fragile state of U.S. tech leadership, such antitrust proposals would play into the hands of foreign rivals. U.S. leadership in four of the five emerging technologies identified by the Biden Administration as a national priority, would be adversely impacted by proposals to restrict digital platforms.
Antitrust policy aiming to limit companies’ size and scope should consider the adverse effects this may have on corporate investments in science and technology which would be detrimental for American technological superiority and national security. We should develop new metrics to show how much corporations contribute to advancing science and technology. Corporations that make significant contributions, especially in fields of national interest, should be treated differently from other big firms that do not make similar contributions.
American economic prosperity and national security rest on its leadership in advancing science, and in applying scientific discoveries to create new goods and services, improving the production and distribution of existing ones. As technologies mature, foreign rivals cut into American firms’ lead, and perhaps even surpass them. This has been the case in consumer electronics, semiconductors, and telecommunications. American policymakers should try to nourish the source of American leadership: its innovation ecosystem, which excels in the discovery of new knowledge as well as the commercial application of that knowledge. In particular, the American innovation ecosystem stands out in developing and deploying general-purpose technologies, such as electricity, polymers, semiconductors, computing, artificial intelligence, and machine learning. Inventing these general-purpose technologies would not have been possible without the massive investments of big corporations in scientific research.
Yet, the United States has lost a substantial amount of corporate research since the 1980s. Corporate research is the source of many breakthrough innovations and cannot be easily substituted by university research and startups. American leadership in emerging technologies critically depends on corporate research and only a few companies continue to invest in research at a meaningful level. The antitrust proposals will impair the ability of these few leading R&D performers to develop new products and enter new markets. The loss of tech companies with scale and scope would reduce U.S. investments in R&D and hurt American economic prosperity and security.
Industrial research is a vital bridge between university discoveries and commercial applications. This is especially so in the case of general-purpose technologies. The decline of industrial research in America is therefore concerning. But although many of the well-known corporate labs — Bell Labs, Xerox PARC, GE Research Labs, GM Research, and DuPont Research, to name a few — are a pale shadow of their former selves, in the ICT sector, some of the slack has been taken up by a few emerging technology market leaders. Notable among them are Alphabet, IBM, and Microsoft, and to a lesser extent, Amazon, Facebook, Intel, and others. These firms have invested both in the creation of new scientific knowledge in artificial intelligence and quantum computing, and in the development of new commercial applications for them. When they succeed, it is likely that this will result in the creation of new products, and in integrating existing products that are united by a shared technology base. Restricting innovators from introducing new products or entering new market segments will undoubtedly reduce their incentive to create the next general-purpose technology, the best way of protecting America’s lead in science and technology.
9 Supplementary Material
HISTORICAL BACKGROUND: THE EVOLVING AMERICAN INNOVATION ECOSYSTEM
The American innovation ecosystem has evolved over time and become more specialized. Universities are focusing on scientific research, startups are experimenting with the commercial applications of university discoveries, and most incumbent producers are withdrawing from research to focus on downstream development. Government support for R&D has diminished as has its role as a leading buyer of advanced technology products. Though more efficient in many ways, the new ecosystem is vulnerable to a fall in the share of basic and applied research in business R&D, and the decline of industrial research. The supplementary material provided below describes how the ecosystem has evolved and the challenges it currently faces. 31
THE RISE OF CORPORATE SCIENCE
By the turn of the 20th century, the United States had become the world’s richest economy and Americans had the highest income per capita by 1914. However, scientifically speaking, American companies were behind and were borrowing cutting-edge technology rather than creating it. 32 The leading American corporations of the 1870s and 1880s largely relied on external inventions. The railroad companies did not invent steam engines or braking systems, nor did Western Union invent the telegraph. Instead, they acquired novel products and ideas from inventors overseas in an active market for technology. Large companies nevertheless established corporate laboratories to evaluate the quality of external inventions, test materials and control quality, and troubleshoot production-related issues. But by World War I, some of the leading American firms recognized they could no longer rely on borrowed technologies or individual inventors. Companies like General Electric (GE) and DuPont began to invest directly in scientific research in electronics and polymer chemistry, respectively. The corporate R&D lab had emerged.
GE illustrates how science-based competition from across the Atlantic drove the establishment of corporate labs in the United States. In the 1890s, GE’s control over electric lighting was based on the carbon- filament high-vacuum incandescent light bulb that had been invented by Thomas Edison in 1879. Shortly after Edison’s invention, German chemist Carl Welsbach came up with a substitute product, the incandescent mantle for gas lamps. Subsequently, German chemist Walther Nernst developed a glower that did not require a vacuum and was 50% more efficient than the carbon-filament high-vacuum incandescent. Patent rights to the Nernst glower were first sold to the German firm AEG, and then to GE’s rival, Westinghouse, in 1894. GE management noticed a flurry of innovative activity that was difficult to control and so approved electrochemist Charles Steinmetz’s proposal to establish the GE Research Laboratory (GERL) in 1900. The technical returns came quickly: in 1906, William Coolidge developed a method for increasing bulb life using tungsten filaments; in 1913, Irving Langmuir invented the inert gas-filled light bulb that became the industry standard. The economic returns of these innovations took longer to materialize, as new machinery had to be developed and then diffused widely.
DuPont is a leading example of the development of the American corporate lab during the interwar period. The company had risen to prominence during World War I because of a massive demand for its explosives products. War-time profits and access to German technology allowed it to grow by expanding into coal-tar chemicals, nitrocellulose lacquers and paints, synthetic fibers and plastics, and gasoline additives (through partnerships with General Motors, Standard Oil, and Dow). DuPont initially diversified from explosives into products that shared a nitrocellulose base, such as lacquers, and semi-synthetic fibers such as rayon. In turn, these products led DuPont to invest in polymer science, which is the basis for many paints, rubbers, plastics, and synthetic fibers. 33 The potential applications of polymer science matched DuPont’s ability to apply them in a variety of products, uniting a disparate range of products. DuPont also had the scale and accumulated manufacturing experience to bring the resulting products to market. 34 Earning private commercial returns from these new products required patience and investment in developing the market. For instance, new synthetic fibers had to be improved so that they could be dyed and developed so that existing textile machines could spin the yarn and weave the fabric. DuPont’s scale and resources were instrumental in bringing together the different elements required to commercialize nylon.
Corporate scientific research was further developed during World War II. Inventions such as mass-produced penicillin, the codebreaking computers at Bletchley Park, radar and even the atomic bomb, were either researched, developed, or refined by American corporate researchers working closely with researchers from federal laboratories and universities. 35 In the wake of war, American corporate science became recognized around the world for its high quality. For example, AT&T’s Bell Labs — home to the transistor, laser, and information theory — produced 14 Nobel Prize winners and five Turing Award recipients, one of the highest honors in computer science. This type of fundamental scientific research performed in corporate labs also spurred follow-on innovation.
Corporations, universities, and the government were key institutions of the post-World War II innovation ecosystem. The government supported most basic research through universities and federal laboratories and with defense procurement financed a significant share of early- stage work on new technologies. Large corporate labs made major scientific contributions, especially when the research was expected to have important commercial applications. They also took important first steps in translating cutting-edge research into prototypes and new products.
THE NEW ECOSYSTEM: DIVISION OF INNOVATIVE LABOR AND THE DECLINE OF INDUSTRIAL RESEARCH
The heyday of the corporate lab lasted from the end of World War II until the 1980s, by which time it was in steady decline. Many leading American firms began to withdraw from performing scientific research, by shutting down their labs, orienting them toward applied activities, or spinning them off completely. One example is AT&T who in 1996 spun off Bell Labs, after seven decades of operations, to form Lucent Technologies. Lucent merged with Alcatel in 2006, and the resulting Alcatel-Lucent company was acquired by Nokia in 2015. The American innovation ecosystem became increasingly specialized, moving toward an economy with a starker division of labor. Universities specialized in research, whilst small startups, often founded by university scientists and financed by venture capital, converted promising scientific discoveries into inventions, and larger, more established firms specialized in product development and commercialization.
The decline of corporate science might erroneously be attributed to a decrease in the usefulness of science for the application of inventions. Yet, patents have continued to cite the science and engineering literature at ever-increasing rates. Moreover, patents that cite science are, on average, of higher quality and more valuable. 36 Indeed, the issue is that scientific knowledge diffuses rapidly. Instead of building on internal research, corporations are now relying on universities, national labs, and other public research institutes. Corporate patents increasingly build on external research. These trends are reflected in the private sector’s declining valuation of research. 37 Therefore, scientific research is not becoming less relevant for innovation; companies are just not as willing to invest in the long slog of “doing” science. This creates a general misconception about the usefulness and overall value of science as firms continue to rely on external science in their inventions.
In summary, ample evidence suggests that over the past several decades the American innovation ecosystem has become more specialized, but also more decentralized and fragmented. Universities and startups have focused on research, and corporations have focused on development while scaling back their investment in scientific research. While this new ecosystem may be more efficient, it also has weaknesses. A vibrant innovation ecosystem requires strong market leaders to complement innovative startups and world class universities.
WHY THE LOSS OF CORPORATE RESEARCH IS A CAUSE FOR CONCERN
In the new specialized ecosystem, scientific research is conducted in universities, while development, engineering, and manufacturing take place in corporations. Specialization means that universities can become better at producing basic research and corporations at developing products. Standard economic theory suggests that specialization is beneficial to both firms and society. However, investment in research has substantial external benefits without which markets are likely to fail to produce an optimal social outcome. For example, Xerox PARC has created large benefits for the American economy through inventions such as the mouse-controlled graphical user interface used by virtually every desktop computer, Ethernet for local area networks, and, of course, the laser printer. 38 Xerox was willing to invest in the Palo Alto Research Center (PARC) in part because PARC offered the possibility of creating new products, such as the laser printer, to fuel the growth of the firm.
The synergy between science and its application finds its natural expression in industrial research. 39 Industrial research projects differ from the investigator-initiated projects conducted in universities and also from those in startups. Compared to university labs, corporate labs conduct research directed towards solving more practical problems. Moreover, unlike small firms that often must scramble for survival, large corporations have the capacity to provide researchers with some slack. Still, commercial necessity is paramount to corporate labs. As Andrew Odlyzko, who spent over two decades at Bell Labs, noted: “It was very important that Bell Labs had a connection to the market … and thereby to real problems. The fact that it wasn’t a tight coupling is what enabled people to work on many long-term problems. But the coupling was there, and so the wild goose chases that are at the heart of really innovative research tended to be less wild, more carefully targeted and less subject to the inertia that is characteristic of university research.” 40 Corporate labs have the potential to integrate the best of both worlds: their research is connected to real problems, so the results are likely to have important industrial applications. The connection is not so strong that the results lie at the most applied end of the spectrum and only provide limited scientific value.
The early impact of corporate science was amplified by the close interactions between corporate labs and other sectors of the U.S. innovation ecosystem. An example is Xerox PARC, the pioneer of modern office technology and arguably the most innovative corporate research lab of the 1970s. But many elements of PARC’s innovations came from outside the organization, most notably from the ARPA-funded Augmentation Research Center (ARC) at the Stanford Research Institute. The ARC had developed bit-mapped screens, the mouse, hypertext, collaborative tools and precursors to the graphical user interface in the mid-1960s, long before the private sector used them. PARC, which hired many ARC researchers, benefited greatly from the early absorption of these technologies. Subsequently, PARC’s innovations spilled over to other organizations. After 24-year-old Steve Jobs visited PARC in 1979, he incorporated many key PARC innovations into the Apple Lisa and the Macintosh. In another example, Charles Simonyi, who had developed the first user-friendly word processor for PARC (the Bravo), left PARC to take a job at Microsoft, where he oversaw the creation of the Microsoft Office suite of applications. The relationship between the different components of the U.S. innovation ecosystem was often symbiotic. University scientists focused on an open-ended search for knowledge, not the commercialization of new products. Xerox PARC scientists improved and integrated ideas coming from universities and research institutions, but their products were intended for experts. Meanwhile, the startup Apple created a computer for the masses. Because each component of the innovation ecosystem had its strengths and weaknesses, it was only by working together that they were able to create the personal computer industry.
Today as much as then, a vibrant U.S. innovation ecosystem needs corporate labs to work in conjunction with world-renowned universities and startups, each playing their part. Corporations have access to specialized resources and expensive equipment, as well as links to manufacturing that can inform and scale useful research. Consequently, corporations can tackle projects of much larger scale and scope than universities. Corporate and university researchers also have different incentives. University researchers are rewarded for being first to make novel discoveries, whereas corporate researchers are rewarded for producing discoveries that are useful to their employers. University research is more likely to be new, but less likely to have commercial use — at least not right away. Corporate researchers are thought to be less likely to rush to print with unreliable discoveries. Similarly, machine learning publications from large firms are cited more often in patents than other machine learning publications.
The close collaboration between science and engineering is another corporate advantage that is harder to replicate in universities or startups. The Google Translate project is a case in point. To implement code written by Quoc Le, one of the leading scientists on the Google Translate project, software engineers converted Le’s code into Google’s newly developed Tensor Flow language, while hardware engineers debugged Google’s proprietary Tensor Processing Units (TPUs), which were custom-built by Google for inference tasks in neural networks. 41 Google has steadily improved these TPU chips, introducing four generations within two years.
The machine translation example also highlights the multi-disciplinary nature of mission-oriented research. The transistor, for instance, would not have been possible without the blend of theoretical prowess and engineering skills available at Bell Labs. Attempts at solid state electronics had been made since the early 1940s by Purdue’s Karl Lark-Horovitz, General Electric, and others. Only Bell Labs, however, had the interdisciplinary team of physicists, metallurgists, and chemists necessary to solve the many theoretical and practical problems associated with developing the transistor. MIT’s Radiation Lab had selected AT&T’s Western Electric subsidiary to manufacture back-voltage rectifiers for radars during World War II and so metallurgists at the firm were able to gain first-hand experience in purifying and doping semiconductors. Bell metallurgist Henry Theurer later developed the method of zone refining in 1951, which processed germanium crystals to impurity levels as low as one part in ten billion. It was also at Bell Labs that Teal and Beuhler’s crystal “pulling” method for fabricating the positive-negative junctions in silicon rods was developed. The key manufacturing process, however, turned out to be “zone refining” developed by W. G. Pfann at Bell Labs. 42 Shockley’s transistor could not have been commercially successful without either one of these in-house achievements in material sciences. Similarly, cross-functional coordination between R&D and manufacturing led to Fairchild Semiconductor’s two major breakthroughs: the planar process and integrated circuits. 43
Important as the development of the solid-state semiconductor was, its application to practical problems required numerous other developments in computer design and storage, and the development of the micro-processor by companies such as IBM, Texas Instruments, and Intel. Computers, initially used for solving ballistic problems for the U.S. Army, were soon being used for accounting and tabulating information by insurance companies and other industrial concerns. In due course, computers drastically reshaped the architecture of entire industries. This exemplifies that, even when some of the important technical details are sorted out, discovering the most useful practical applications remains an important task. In the American innovation ecosystem, this task is increasingly being left to entrepreneurial startups.
Startups do not tend to contribute to the production of science. Nevertheless, startup inventions make heavy use of public science, as evidenced by the high number of citations their patents make to scientific publications authored by university researchers. This is indicative of the important role of startups in commercializing science produced by other actors in the innovation ecosystem, most notably universities, but in some cases, also corporate labs. 44 The commercial application of a technological breakthrough faces two types of uncertainty. Technical uncertainty focuses on whether a given technical objective can be achieved using the proposed approach, such as whether antibiotics may help with stomach ulcers. Commercial uncertainty refers to the challenges of accurately assessing demand for the proposed product and the likely costs of scaling up and servicing the market. Juggling both types of uncertainty may prove prohibitive for startups, which explains why startups in the physical sciences have received limited private sector funding.
The energy sector offers a case in point. Commercial success in renewable energy, for example, may require innovations in electronics, materials, chemistry, and, more importantly, changes in regulations for implementation. Thermionic energy generation is a method that directly converts heat to electricity, promising significant improvements on mechanical heat engines. Originally explored in the 1960s for powering satellites, the technology was initially turned down by investors. Only recently, once the microfabrication tools required to create prototypes became easily available, did it attract investment. But beyond the technical challenges, adoption of inventions in the energy sector also requires a change in existing technical infrastructure, consumer behavior, and government regulation — all of which compounded market uncertainty. Clean energy innovations in wind and solar, for instance, depend on the development of grid energy storage technologies, since output fluctuates with weather patterns. But advances in energy storage, such as battery technology, depend on downstream market demand. Owing to these risks, VC funding in battery technology startups only started in earnest in the 2000s, after the automotive sector began adopting hybrid and fully electric vehicles, charging infrastructure was developed, and customers demanded electric vehicles. Large corporations can often better manage these commercial and technical uncertainties because they have experience moving products from labs to markets, and because they, or their partners, can be a source of demand.
In summary, the loss of corporate research is a cause for concern. While a market-driven specialized innovation ecosystem has benefits to both firms and society, it is unlikely to produce an optimal level investment in research due to the large external benefits it creates. Moreover, corporate research is unique. It cannot be replenished by research performed by universities or startups alone.
CASE STUDIES: GENERAL PURPOSE TECHNOLOGIES IN THE ICT SECTOR
SEMICONDUCTORS
The microprocessor that powered desktop computers, servers, and routers, the very basis for the IT evolution and the Internet and digital platforms, is a general-purpose technology by design. It is not customized for specific applications — the same microprocessor might be used to send emails, browse the Internet, handle payroll, route traffic on the Internet, among the myriads of tasks that computers are used for. As noted earlier, both the science and the technology underlying microprocessor design and production were essentially American. Today, there are many claimants for that honor.
The COVID induced shortage of semiconductors has made it painfully clear that semiconductors are not principally manufactured in America. Data from the Semiconductor Industry Association indicate that in 2020, US had only 12% of global chip manufacturing, compared to 22% for Taiwan, 21% for S. Korea, 15% each for Japan and China and Europe 8%. By 2030, the U.S. share is predicted to fall to 10%, and China’s share is predicted to rise to 24%. 45
And even the high value microprocessor segment was, until recently, dominated by American corporations such as Intel and AMD, has since been overshadowed by the rise of chips designed for mobile applications, dominated by ARM, a British company. Increasingly, large users such as Apple, Google, and Tesla design their own chips, customized to their needs, albeit still produced by companies from Taiwan (TSMC) or South Korea (Samsung). And though the basic innovations in fabricating semiconductors and integrating them were pioneered by American firms such as Texas Instruments, ATT (Bell Labs), and Intel, the key equipment for fabricating the current generation of five nanometer semiconductors, namely Extreme Ultraviolet Lithography, are provided by ASML, a Dutch firm, whose current market value exceeds that of Intel.
An examination of patent and scientific publication data reveals the growing dominance of foreign companies. Between 2011 and 2016, American firms (IBM, AT&T, Intel, Micron, Microsoft, Alphabet) accounted for about a quarter of scientific publications, and nearly two thirds of the high-impact scientific publications in semiconductors. But between 2017-2019, American firms accounted for only half of the high-impact scientific publications. Samsung alone published nearly as many high-impact publications (12) as all American firms put together (15). The relative decline in its scientific prowess in semiconductors is also reflected in patent data. Samsung is also the patenting leader. In 2019, it was granted 2,272 patents in the United States, more than double the nearest American IBM, with 924. TSMC, another leading contract fabricator of semiconductors received 2196 patents, followed by Toshiba with 954 patents. The rise of Samsung as a technology leader is likely to accelerate following its announcement in August 2021 that it plans to invest $206 billion over a three-year period on expanding its manufacturing capacity technological capabilities in the fields of semiconductors, biopharmaceutical, robotics industries, and artificial intelligence. Samsung has already started the construction of a new manufacturing line in Pyeongtaek, Korea to be completed by end of 2022. This production line will be the hub for cutting-edge innovations and capable of manufacturing 5nm logic semiconductors based on EUV technology and 14nm DRAM.
Yet not all is lost in the American front. On May 6, 2021, IBM announced it has successfully created the world’s first 2nm node chip, which saves 75% more energy and improves the performance of overall chip by 50%. IBM claims that this technology can fit 50 billion transistors on a chip of a size of a fingernail. This accomplishment is especially impressive because IBM does not own a foundry (instead of owning a fab, IBM signed a 10-year contract with Global Foundries). American universities are also making substantial scientific discoveries that are likely to create new technological opportunities in the near future. For example, a team of electrical and computer engineers from Purdue University discovered how to combine silicon and ferroelectric material to produce the ferroelectric semiconductor field-effect transistor. This discovery makes chips much denser, energy efficient and ultimately small enough for building circuits that can mimic networks in the human brain. At the same time, American startups continue to bridge the gap between universities and big companies by commercializing new technologies at the scientific frontier. An example is Gallium Nitride Semiconductors (GaN). GaN is a wide band gap semiconductor material that is manufactured by taking a standard Silicon wafer and growing a thin layer of GaN on top of Aluminum Nitride using metal organic chemical vapor deposition (MOCVD) process. Due its greater efficiency and reliability at high voltage and high temperatures GaN has applications in a wide range of industries including telecommunications, automotive, renewables, health care and defense. For instance, GaN’s higher switching frequency makes it a perfect semiconductor for 5G technologies. The global market for GaN semiconductors is $20 billion in 2021 and is expected to grow. Currently, GaN semiconductors are manufactured by small firms, mostly American: Qorvo (U.S.), Cree (U.S.), MACOM (U.S.), Infineon Technology (Germany), and GaN System (Canada).
To recap, the overview of the semiconductors ecosystem offers two key takeaways. First, American leadership in high-tech industries cannot be taken for granted. As technology matures, foreign rivals, sometimes supported by their national governments, will develop into credible challengers. The second takeaway is that while America has clearly lost its edge in science, technology and manufacturing, some sectors of its ecosystem — universities and startups — provide reasons for optimism. Yet, this optimism is unlikely to materialize into significant economic benefits unless investment by big American firms will match that of their foreign rivals.
ARTIFICIAL INTELLIGENCE
Artificial Intelligence (AI) research has become a recent prime example of the importance of corporate science even in an innovation ecosystem that includes strong university science. Though universities are naturally producing a large share of scientific advances in artificial intelligence, corporate research was particularly important in the emerging phase of this GPT. The share of corporate publications in such top AI journals as the International Conference on Machine Learning (ICML) have tripled between 2004 and 2016. Between 2011 and 2015, IBM, Microsoft, and Google together produced more high impact publications than MIT(!). 46 Many companies are actively researching and publishing in AI, including Intel, Huawei, Samsung, Nokia Bell Labs, Facebook, NTT, along with Alibaba, Baidu, Apple, and LG. 47 48
Corporate labs have the scale that startups or university labs lack. Google acquired DeepMind in 2014. Recently, DeepMind published a landmark paper with an algorithm for predicting the three-dimensional structure of proteins, along with the predicted structure of 35,000 proteins. 49 This accomplishment required substantial investment. In 2014, DeepMind had 400 employees; today, it employs more than 1,000. Over this period, Google has reportedly invested nearly $2 billion in the company in the form of accumulated losses.
In addition to financial resources, Google has access to fundamental complementary assets that raise the strategic value of AI algorithms. For example, Google uses the JFT-300M dataset, which has more than 375 million labels for 300 million images. Compare that to Stanford’s Imagenet dataset, one of the largest datasets made publicly available by a university, which contains only around 1 million images. More data means significant performance improvements in AI. 50 Beyond sizable datasets, applying AI research requires interdisciplinary teams. Domain specialists — such as linguists in the case of machine translation — define the problem to be solved and assess performance. Statisticians design the algorithms, theorize on their error bounds, and optimize routines. Meanwhile, computer scientists search for efficiency gains in implementing those algorithms. Not surprisingly, the “Google Translate” paper has 31 co-authors, many of whom are leading researchers in their respective fields.
This multidisciplinary approach distinguishes university research from industry AI research. Between 2011 and 2018, research publications by large firms featured, on average, one more co-author (4.3) than other publications (3.4). 51 The top AI publishers, including Microsoft, Google, IBM, Yahoo, Toyota, Baidu, NEC Corporation, Facebook, Adobe, and LinkedIn, 52 make up 10% of the papers published with fewer than 11 authors (2,168 out of 20,989), but 28% of the papers published with more than 11 authors (22 out of 79). High-quality papers show the same difference in team size. Among the Machine Learning Conference papers in the top decile by citations received, corporate publications have 4.4 authors, while non-firm publications have 3.6. This pattern holds also for the top 1% of cited publications. Firm publications involve more co-authors than non-firm publications, with 4.4 co-authors compared to 3.6, respectively.
Scientific research prowess is also reflected in inventive outcomes. There are 27 American firms among the top 50 AI patenting organizations between 2006 and 2020, accounting for nearly 70% of the patents filed by the top 50 patentees. The top 15 patenting organizations are IBM (with 1114 patents), Microsoft (458), Capital One (359), Facebook (245), NEC (209), Google (206), Cisco (191), Oracle (189), Accenture (189), Fujitsu (185), Bank of America (172), Amazon (171), Intel (153), Samsung (141), and LG (123). The other noteworthy aspect, apart from the dominance of American firms, is the variety of industries represented, from computer hardware and systems, software, semiconductors, digital platforms, and networking equipment to R&D and technical services, financial services, online retailing, and electronics. This is the hallmark of a general-purpose technology.
QUANTUM COMPUTING
Large corporate labs’ recent contributions to high tech innovation can best be seen in the rapidly evolving quantum computing industry. Quantum computing is a scientific and technological field that exploits such quantum phenomena as superposition and entanglement to exponentially speed up computations relative to conventional computers. 53 As a result, quantum computers have the potential to revolutionize such diverse industries as cybersecurity, manufacturing and logistics, financial services, artificial intelligence, and chemistry. For example, by enabling scientists to simulate chemical reactions much more precisely than classical computers can, quantum computing could lead to important innovations in pharmaceutical research, materials science, and battery technology.
Though government funded university research has undoubtedly been crucial, quantum computing would not have been possible without sustained investments by corporations. 54 Even before the industry began, in the 1940s, AT&T’s Bell Labs supported Claude Shannon’s work on information theory, as the company stood to benefit the most from a more efficient communication network. Shannon’s research eventually laid the foundation for the modern digital age, influencing not only communication technology, but also probability and statistics, cryptography, and computer science.
Early work in quantum computing was largely theoretical — so called “basic research” centering around the mathematical foundations of the field — and mostly carried out by university scientists. Yet, corporate scientists also made fundamental scientific contributions. In 1984, IBM Research’s Charles Bennett discovered the first quantum cryptography protocol. In the 1990s, he also developed his four laws of quantum information. Today, Bennett is regarded as a founding father of modern quantum information theory. Another fundamental contribution was made in 1994 by Bell Labs’ Peter Shor, who proposed a quantum algorithm that could calculate the prime factors of a large number vastly more efficiently than a classical computer. Because factoring large numbers is the basis for most encryption schemes, Shor’s algorithm sparked tremendous interest in quantum computing.
The main technical challenge to scaling and commercializing quantum computing is error correction. Unlike their logical counterparts, current physical qubits do not hold their state long enough — due to interactions with the environment, such as touching another particle or being exposed to radiation — to enable accurate calculations. The current solution for minimizing errors is to cool the hardware to near absolute zero in a large vacuum chamber. This substantially raises the cost of building and operating a quantum computer, as well as the physical size of current systems. The most popular implementation today uses superconducting qubits built with superconducting small electrical loops operating at ultra-low temperatures. This method is used by large firms like IBM and Google, but also by smaller firms like Rigetti Computing. Conversely, academic offshoot IonQ uses a “trapped ions” implementation that confines and suspends charged ytterbium atoms in vacuum using electromagnetic fields. Both Rigetti and IonQ make their hardware available through Amazon Braket, underscoring how important large-firm collaborations are to the industry’s smaller players.
Leading U.S. corporations — including IBM, Intel, Microsoft, Amazon, and Google — all have invested in quantum computing. For example, IBM collaborates with external partners to develop such applications as the Q Network, a cloud computing service that allows research institutions, startups, and large companies to use IBM’s more than 20 different quantum processors. Firms such as Daimler, Samsung, and Goldman Sachs are currently exploring the technology with IBM. Like IBM, Amazon offers Braket, a managed service operating on the AWS Cloud designed to speed up scientific research and software development for quantum computing. But unlike IBM, the resulting programs can be deployed on quantum hardware built by smaller third-party companies — including D-Wave, IonQ, and Rigetti — not on Amazon’s hardware. The company is also actively performing scientific research into quantum error-correction, an essential step in building useful quantum chips that are stable (unlike current qubits). Meanwhile, Microsoft is developing a “topological” quantum computer where information is stored not in any single particle, but in the collective behavior of an entire wire. 55 In August 2020, Google researchers further demonstrated the potential of quantum technology to be used in real-world applications by successfully simulating a simple chemical reaction. In May 2021, Google unveiled its new Quantum AI campus that brings the company’s quantum computing activities under one roof with the goal of building an error-corrected, useful quantum computer in the next decade.
It is believed that the Chinese government makes multi-billion-dollar investments to complement efforts by corporations such as Huawei, Baidu, and Alibaba. 56 For example, in 2016, China launched the first quantum satellite that transmitted pairs of entangled photons to base stations located hundreds of miles apart. The technology enabled the first conference call secured through quantum key distribution. More recently, in December 2020, Chinese researchers also claimed quantum supremacy, building the most powerful photonic quantum computer to date. Private Chinese firms have also invested in quantum computing research. Between 2005 and 2020, Alibaba, Baidu, and Huawei produced a total of 40 scientific papers on quantum computing. These are only a fraction of those produced by IBM, which alone has published 291 papers, or Microsoft, which has published 249 papers.
Due to their size, large corporations are well positioned to invest the resources needed to overcome technical challenges. Due to their scope, they are also well positioned to overcome uncertainty regarding their future ability to commercialize quantum computing technologies in private markets. It is not surprising, then, to see most tech leaders participating in both upstream scientific research and downstream technology development for quantum computing. Meanwhile, smaller firms participate to benefit from selling their specialized products and services to large corporations. In short, quantum computing shows how well the innovation ecosystem in America works when the market leaders also invest in scientific research.
Among the top 50 patentees in quantum computing over the period 2013-2020, thirty are American, and account for nearly 80% of the patents by the top 50 patentees. IBM leads the pack with 233 patents, followed by Microsoft (66), Intel (63), Northrop Grumman (42), and Alphabet (41). Three startups, Equal1 Labs, D-Wave, and Righetti, from Ireland, Canada, and the United States, respectively, are in the middle of the pack, along with Toshiba from Japan. Two American universities, MIT and Yale, are tied for tenth place with 14 patents each. Other prominent patentees include Accenture, Honeywell, Bank of America, Caltech, and Amazon, consistent with quantum computing being a general-purpose technology with diverse applications. The presence of Yale, MIT, and Caltech among the list of top patentees also shows that quantum computing advances are closely related to scientific research in quantum computing. Many of the tech firms that lead in quantum computing patenting are also prominent in quantum computing research. For example, IBM, Microsoft, and Google together produced 52 high-impact scientific publications between 2005 and 2020, more than MIT, which produced 47 high-impact scientific papers.
Despite its impressive achievements, quantum computing as an industry is facing substantial challenges. Quantum computing technology is currently noisy, experimental, and small in scale. While communication applications such as quantum key distribution and quantum encryption are in use today, quantum computers are yet to supply real-world benefits or diffuse in the economy as a GPT. However, such breakthroughs as showing quantum supremacy have demonstrated quantum computing’s potential impact on society. U.S. companies have introduced roadmaps for the next few years. For example, Google aims to produce a commercial-grade quantum computer by 2029. 57 China has prioritized investing huge amounts of resources in quantum research and development. 58 To retain its comparative advantage in quantum computing science, the United States must encourage close interactions between the different components of the innovation ecosystem. And that includes large corporations’ continued involvement in scientific research, as their large scale and scope uniquely positions them to provide the resources and have the market incentives needed to speed up technology development.
References
“2021 AI City Challenge,” AI City Challenge, accessed May 3, 2022, https://www.aicitychallenge.org/2021-ai-city/.
Ashish Arora, Sharon Belenzon, and Andrea Patacconi, “The Decline of Science in Corporate R&D,” Strategic Management Journal 39, no. 1 (2018): pp. 3-32, https://doi.org/10.1002/smj.2693.
Ashish Arora, Sharon Belenzon, and Lia Sheer, “Knowledge Spillovers and Corporate Investment in Scientific Research,” American Economic Review 111, no. 3 (2021): pp. 871-898, https://doi.org/10.1257/aer.20171742.
Ashish Arora et al., “The Changing Structure of American Innovation: Some Cautionary Remarks for Economic Growth,” Innovation Policy and the Economy 20 (2020): pp. 39-93, https://doi.org/10.1086/705638.
Sara Castellanos, “Google Aims for Commercial-Grade Quantum Computer by 2029,” The Wall Street Journal, May 18, 2021, https://www.wsj.com/articles/google-aims-for-commercial-grade-quantum-computer-by-2029-11621359156.
Daniel Garisto, “China Is Pulling Ahead in Global Quantum Race, New Studies Suggest,” Scientific American, July 15, 2021, https://www.scientificamerican.com/article/china-is-pulling-ahead-in-global-quantum-race-new-studies-suggest/.
Philipp Hartmann and Joachim Henkel, “The Rise of Corporate Science in AI: Data as a Strategic Resource,” Academy of Management Discoveries 6, no. 3 (October 29, 2020), https://doi.org/10.5465/amd.2019.0043.
Matthew Heller, “Google Waives $1.5B in Deepmind Startup Debt,” CFO Magazine, December 17, 2020, https://www.cfo.com/accounting-tax/2020/12/google-waives-1-5b-in-deepmind-startup-debt-4430/.
Dean, J., Corrado, G., Monga, R., Chen, K., Devin, M., Mao, M., Ranzato, M.A., Senior, A., Tucker, P., Yang, K. and Le, Q., 2012. Large scale distributed deep networks. Advances in neural information processing systems, 25. Marx, M. (2019, February). Patent Citations to Science. Retrieved 2019-03-03, from https://www.openicpsr.org/openicpsr/project/108362/version/V12/view/ (type: dataset) doi: 10.3886/ E108362V12
Derek Hill, “Production and Trade of Knowledge- and Technology-Intensive Industries: Figure 6-D,” Science and Engineering Indicators, January 2020, https://ncses.nsf.gov/pubs/nsb20205/production-patterns-and-trends-of-knowledge-and-technology-intensive-industries#figureCtr990.
Derek Hill, “Table S-19: Exports and Imports of High R&D Intensive Products, by Region, Country, or Economy: 2003–18,” Science and Engineering Indicators, January 2020, https://ncses.nsf.gov/pubs/nsb20205/data.
“Industrial Research Laboratories of the United States, Including Consulting Research Laboratories,” National Research Council of the National Academy of Sciences (Washington, D.C., 1980).
“Industrial Research Laboratories of the United States, Including Consulting Research Laboratories,” National Research Council of the National Academy of Sciences (Washington, D.C., 1998).
“Intel to Keep Its Number One Semiconductor Supplier Ranking in 2020,” IC Insights, November 23, 2020, https://www.icinsights.com/news/bulletins/Intel-To-Keep-Its-Number-One-Semiconductor-Supplier-Ranking-In-2020.
John Jumper et al., “Highly Accurate Protein Structure Prediction with Alphafold,” Nature 596, no. 7873 (2021): pp. 583-589, https://doi.org/10.1038/s41586-021-03819-2.
Joshua Lev, Krieger, Martin Watzinger, and Monika Schnitzer. “Standing on the Shoulders of Science.” Harvard Business School Working Paper, no. 21–128 (June 2021). https://www.hbs.edu/faculty/Pages/item.aspx?num=60344.
“Main Science and Technology Indicators,” Organisation for Economic Co-operation and Development (OECD), latest data from March 2022, https://www.oecd.org/sti/msti.htm.
National Science Board, Science Indicators 1980: Report of the National Science Board, 1981 (Washington, D.C.: U.S. Government Printing Office, 1981).
“Quantum Computer Datasheet,” Quantum AI Google, May 14, 2021, https://quantumai.google/hardware/datasheet/weber.pdf.
Kendall A. Smith, “Louis Pasteur, the Father of Immunology?,” Frontiers in Immunology 3, no. 68 (2012), https://doi.org/10.3389/fimmu.2012.00068.
Chen Sun et al., “Revisiting Unreasonable Effectiveness of Data in Deep Learning Era,” 2017 IEEE International Conference on Computer Vision, 2017, pp. 843-852, https://doi.org/10.1109/iccv.2017.97.
“Total Factor Productivity at Constant National Prices for China,” University of Groningen and University of California, Davis, retrieved from FRED, Federal Reserve Bank of St. Louis, last updated November 8, 2021, https://fred.stlouisfed.org/series/RTFPNACNA632NRUG.
Alan Weissberger, “Dell’Oro: Telecommunication Equipment Market 1q20 to 3q20 +China’s New 5G Base Stations,” Technology Blog, December 3, 2020, https://techblog.comsoc.org/2020/12/03/delloro-telecommunication-equipment-market-1q20-to-3q20-chinas-new-5g-base-stations/.
“Who Leads the 5G Patent Race November 2021?,” IPlytics, November 3, 2021, https://www.iplytics.com/report/5g-patent-race-november-2021/.
“Winning the Future,” Semiconductor Industry Association, April 2019, https://www.semiconductors.org/wp-content/uploads/2019/04/SIA_Winning-the-Future_Refresh_FINAL1.pdf.
“WIPO IP Statistics Data Center,” WIPO, last updated November 2021, https://www3.wipo.int/ipstats/.
Yonghui Wu et al., “Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation,” October 2016, https://doi.org/https://doi.org/10.48550/arXiv.1609.08144.
Remco Zwetsloot et al., “China is Fast Outpacing U.S. STEM PhD Growth,” (Georgetown University Center for Security and Emerging Technology, August 2021), https://doi.org/10.51593/20210018.
[1] “Main Science and Technology Indicators,” Organisation for Economic Co-operation and Development (OECD), latest data from March 2022, https://www.oecd.org/sti/msti.htm.
[2] “Main Science and Technology Indicators.”
[3] “Total Factor Productivity at Constant National Prices for China,” University of Groningen and University of California, Davis, retrieved from FRED, Federal Reserve Bank of St. Louis, last updated November 8, 2021, https://fred.stlouisfed.org/series/RTFPNACNA632NRUG.
[4] “Total Factor Productivity.”
[5] Remco Zwetsloot et al., “China is Fast Outpacing U.S. STEM PhD Growth,” (Georgetown University Center for Security and Emerging Technology, August 2021), https://doi.org/10.51593/20210018.
[6] NSF Science and Engineering Indicators 2020, Table S8-62, https://ncses.nsf.gov/pubs/nsb20204/assets/supplemental-tables/tables/nsb20204-tabs08-062.xlsx.
[7] NSF Science and Engineering Indicators 2020, Figure 22, https://ncses.nsf.gov/pubs/nsb20201/data.
[8] Science-based ICT patents are those in the top quartile of the number of citations made to non-patent literature, relative to other ICT patents granted in the same year.
[9] “WIPO IP Statistics Data Center,” WIPO, last updated November 2021, https://www3.wipo.int/ipstats/.
[10] Derek Hill, “Production and Trade of Knowledge- and Technology-Intensive Industries: Figure 6-D,” Science and Engineering Indicators, January 2020, https://ncses.nsf.gov/pubs/nsb20205/production-patterns-and-trends-of-knowledge-and-technology-intensive-industries#figureCtr990.
[11] “Winning the Future,” Semiconductor Industry Association, April 2019, https://www.semiconductors.org/wp-content/uploads/2019/04/SIA_Winning-the-Future_Refresh_FINAL1.pdf.
[12] High R&D intensive products include aircraft, pharmaceuticals, and computer, electronic, and optical products classified
by the OECD.
[13] Derek Hill, “Table S-19: Exports and Imports of High R&D Intensive Products, by Region, Country, or Economy: 2003–18,” Science and Engineering Indicators, January 2020, https://ncses.nsf.gov/pubs/nsb20205/data.
[14] “Critical and emerging technologies” included in the White House’s multi-agency R&D priorities for formulating fiscal year 2023 budget submissions to the Office of Management and Budget, https://www.whitehouse.gov/wp-content/uploads/2021/07/M-21-32-Multi-Agency-Research-and-Development-Prioirties-for-FY-2023-Budget-.pdf.
[15] “2021 AI City Challenge,” AI City Challenge, accessed May 3, 2022, https://www.aicitychallenge.org/2021-ai-city/.
[16] Microsoft Academic Graph, https://www.microsoft.com/en-us/research/project/microsoft-academic-graph/.
[17] U.S. Patent and Trademark Office, https://patentsview.org/download/data-download-tables. However, patents may be peripheral for quantum computing as they relate mostly to component technologies.
[18] “Quantum Computer Datasheet,” Quantum AI Google, May 14, 2021, https://quantumai.google/hardware/datasheet/weber.pdf.
[19] “Who Leads the 5G Patent Race November 2021?,” IPlytics, November 3, 2021, https://www.iplytics.com/report/5g-patent-race-november-2021/.
[20] Alan Weissberger, “Dell’Oro: Telecommunication Equipment Market 1q20 to 3q20 +China’s New 5G Base Stations,” Technology Blog, December 3, 2020, https://techblog.comsoc.org/2020/12/03/delloro-telecommunication-equipment-market-1q20-to-3q20-chinas-new-5g-base-stations/.
[21] “Winning the Future.”
[22] “Intel to Keep Its Number One Semiconductor Supplier Ranking in 2020,” IC Insights, November 23, 2020, https://www.icinsights.com/news/bulletins/Intel-To-Keep-Its-Number-One-Semiconductor-Supplier-Ranking-In-2020.
[23] Sources: National Center for Science and Engineering Statistics and U.S. Census Bureau, Business Research and Development
Survey series available at https://www.nsf.gov/statistics/srvyberd/. National Bureau of Statistics of China, Basic Statistics on R&D Activities of Industrial Enterprises Above Designated Size by Registration Status, from National Data series available at https://data.stats.gov.cn/english/easyquery.htm?cn=C01.
[24] Data are based on annual reports by performers in the National Center for Science and Engineering Statistics’ annual RD expenditure surveys. National Center for Science and Engineering Statistics, National Patterns of RD Resources: 2018-19 Data Update available at https://ncses.nsf.gov/pubs/nsf21325.
[25] An early commercial application of computers came from an unlikely source. J Lyons & Co, an English company running tea shops, pioneered the first computer for running routine office tasks in 1947. It developed the LEO, which ran its first business application in 1951, to calculate the cost of ingredients that went into bread and cakes. Lyons diversified into manufacturing and selling computers to others, forming LEO Computer Ltd in 1954. https://en.wikipedia.org/wiki/LEO_(computer).
[26] Matthew Heller, “Google Waives $1.5B in Deepmind Startup Debt,” CFO Magazine, December 17, 2020, https://www.cfo.com/accounting-tax/2020/12/google-waives-1-5b-in-deepmind-startup-debt-4430/.
[27] “Industrial Research Laboratories of the United States, Including Consulting Research Laboratories,” National Research Council of the National Academy of Sciences (Washington, D.C., 1980).
[28] “Industrial Research Laboratories of the United States, Including Consulting Research Laboratories,” National Research Council of the National Academy of Sciences (Washington, D.C., 1998).
[29] Ashish Arora, Sharon Belenzon, and Lia Sheer, “Knowledge Spillovers and Corporate Investment in Scientific Research,” American Economic Review 111, no. 3 (2021): pp. 871-898, https://doi.org/10.1257/aer.20171742.
[30] For further analyses, see Arora, Belenzon, and Patacconi (2018).
[31] This draws upon Arora et al. (2020), which also provides more detailed references and sources.
[32] Andrew Carnegie, for instance, made a fortune by implementing — not inventing — the Bessemer process that had been imported from England.
[33] For example, polyester terephthalate (PET) can be used to produce textiles for shirts, staple for carpets, but also plastic for bottles. Nylon polymers are the basis for lingerie, but also parachute cord and engineered plastic parts.
[34] For example, when a key input for nylon was not commercially available, DuPont used its explosives and fertilizer experience to develop and produce it in-house.
[35] Wartime contractors were selected based not only on their R&D capabilities, but also on their ability to produce and distribute the resulting products and services. This policy largely excluded smaller companies from participating in the R&D process.
[36] Krieger, Shnitzer and Waldinger (2021) find that patents that build directly on science are on average 26% more valuable than patents in the same technology class that are disconnected from science. Patents closer to science are also more likely to be in the tails of the value distribution (i.e., greater risk and greater reward), and are more novel.
[37] Estimates from Arora et al. (2018) suggest that the stock market value associated with one publication by a firm with an established research program fell from $900,000 in 1980-1990 to $295,000 in 1990-2000, representing a drop of 67%. Between 1980 and 2006, the market value attributable to the stock of scientific knowledge among an established firm’s intangible assets declined by about 30%.
[38] The laser printer was a blockbuster product for Xerox, but the company only appropriated a small fraction of the benefits created by many of its other innovations.
[39] The 1981 report of the National Science Board begins by pointing out that science and its application are closely intertwined. Significant new discoveries are often made in the course of solving specific problems. Louis Pasteur, studying how to prevent wine from spoiling, discovered the germ theory of fermentation and the widely used technique of Pasteurization to prevent spoilage in such foods as milk, cheese, and beer. In addition to being an extremely valuable industrial innovation, this discovery also led to the modern science of bacteriology, immunology, and vaccines (Smith, 2012).
[40] Letter to the Wall Street Journal, available at http://www.dtc.umn.edu/odlyzko/misc/wsj-bell-labs-20120326.
[41] TPUs are custom Application Specific Integrated Circuits (ASIC) designed for deep neural networks. The first TPUs were deployed in Google data centers in 2015 and performed up to 26 times faster than existing GPUs. https://cloud.google.com/blog/products/gcp/an-in-depth-look-at-googles-first-tensor-processing-unit-tpu
[42] National Science Board, Science Indicators 1980: Report of the National Science Board, 1981 (Washington, D.C.: U.S. Government Printing Office, 1981).
[43] As E.B. Craft of Bell Labs put it: “Perhaps the outstanding characteristic of this organization, the one that sets it apart a little from others, is its conduct of research and development by a group method of attack… the result is the necessity of a high degree of specialization. So in all of these technical departments we have specialist, chemists, metallurgists, physicists, engineers, statisticians, mathematicians, men that are trained and skilled in their particular branches of science and engineering. Their activities are so coordinated by means of this organization, that their best brains can be brought to bear upon any specific problem… When a problem is put up to the Labs for solution, it is divided into the elements and each element is assigned to that group of specialists who know the most about that particular field but they all cooperate and make their contribution to the solution of the problem as a whole” (America by Design, page 119).
[44] For example, in 1979 Robert Metcalfe founded 3Com to commercialize the Ethernet-based local area networking technology he and others had developed while working at Xerox PARC.
[45] “Winning the Future.”
[46] Philipp Hartmann and Joachim Henkel, “The Rise of Corporate Science in AI: Data as a Strategic Resource,” Academy of Management Discoveries 6, no. 3 (October 29, 2020), https://doi.org/10.5465/amd.2019.0043.
[47] Google has published landmark papers such as the “Cat Paper” and the “Google Translate Paper” that validated the effectiveness of new algorithms such as LSTM (Long-Short Term Memory) for image recognition and language translation respectively.
[48] Yonghui Wu et al., “Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation,” October 2016, https://doi.org/https://doi.org/10.48550/arXiv.1609.08144.
[49] John Jumper et al., “Highly Accurate Protein Structure Prediction with Alphafold,” Nature 596, no. 7873 (2021): pp. 583-589, https://doi.org/10.1038/s41586-021-03819-2.
[50] Chen Sun et al., “Revisiting Unreasonable Effectiveness of Data in Deep Learning Era,” 2017 IEEE International Conference on Computer Vision, 2017, pp. 843-852, https://doi.org/10.1109/iccv.2017.97.
[51] This was found by looking at data from Marx (2019) and examining the average number of coauthors in the five leading machine learning conferences in the given time period (Hartmann & Henkel, 2019). The five conferences are Knowledge Discovery and Data Mining (KDD), the Association for the Advancement of Artificial Intelligence (AAAI), the International Conference on Machine Learning (ICML), the International Joint Conferences on Artificial Intelligence (IJCAI), and the Conference on Neural Information Processing Systems (NIPS).
[52] Hartmann and Henkel, “The Rise of Corporate Science.”
[53] Conventional computers manipulate individual bits, which store information as binary 0 or 1 states. Quantum computers use quantum bits, or qubits, which can represent multiple combinations of 0 and 1 at the same time. This capability, called superposition, has the potential to magnify computing power for specific algorithms. Conversely, quantum entanglement is the ability of a pair of qubits to exist in a single quantum state. Changing the state of one qubit instantly changes the state of the other qubit in a predictable way, even if they are separated by long distances. This property allows quantum computers to exponentially increase computational ability with the addition of just a few extra qubits.
[54] Consulting firm Qureca estimates that governments worldwide invested approximately $22 billion in quantum research and technologies in 2020.
[55] In 2019, Google researchers reported the first experimental realization of “quantum supremacy,” when in about 200 seconds, Google’s quantum computers performed calculations that would have taken a state-of-the-art conventional supercomputer about 10,000 years to carry out. IBM later claimed this was not entirely correct.
[56] President Xi Jinping has said that “quantum science is a matter of national concern.”
[57] Sara Castellanos, “Google Aims for Commercial-Grade Quantum Computer by 2029,” The Wall Street Journal, May 18, 2021, https://www.wsj.com/articles/google-aims-for-commercial-grade-quantum-computer-by-2029-11621359156.
[58] Daniel Garisto, “China Is Pulling Ahead in Global Quantum Race, New Studies Suggest,” Scientific American, July 15, 2021, https://www.scientificamerican.com/article/china-is-pulling-ahead-in-global-quantum-race-new-studies-suggest/.
Past Event: How the U.S. Can Better Compete Globally on R&DBuilding a Stronger (More Complex) U.S. Manufacturing SectorPast Event: How Better Statistics Lead to Better Policy in a Changing WorldGeothermal Everywhere: A New Path for American Renewable Energy LeadershipSlide Deck–American Innovation Under Threat: Restrictive Legislation and Global Competition