The Electronic Numerical Integrator and Computer (ENIAC), which became operational in 1945, was by many accounts the world’s first programmable computer. The machine was developed at the University of Pennsylvania with funding from the U.S. Army. Credit: U.S. Army
Attempts to replicate a complex global semiconductor manufacturing ecosystem on American soil are inefficient and costly. Instead, U.S. policymakers should seek to onshore only a minimum viable capacity of chip production as an insurance policy against emerging security risks.
When Mark Liu, chairman of the world’s largest chipmaker, got on a call with investors and analysts in January, it was to deliver a message that would have ripple effects not just for his company, Taiwan Semiconductor Manufacturing Co., but also for U.S. policy makers.
TSMC, as the company is known, had already delayed the opening of its new Arizona semiconductor fabrication plant to 2026. But Liu, under increasing pressure from the Biden administration to deliver on the multibillion-dollar plan, told everyone who dialed in to the earnings call that he needed even more time to get the U.S. production up and running.
According to the TSMC chairman, the Arizona factory would now only manage to be ready by 2027 or 2028. “There will be a gap,” he said on the call.2
Liu’s statement was only the latest setback for the United States’ semiconductor strategy. That strategy was most representatively spelled out by Secretary of Commerce Gina Raimondo in September 2023, when she remarked that “none of [the most advanced chips] are made in America right now, which is why I’ve got to work like crazy every day to bring that manufacturing home as fast as we can.”3
Which Chips?
Working out the best way to safeguard the semiconductor supply chain requires first discerning the most vulnerable segment of that supply chain. The production of advanced semiconductors (requiring <10nm nodes to produce), contrasted with that of less-advanced semiconductors (requiring >10nm nodes to produce), is particularly vulnerable for two reasons.
Firstly, advanced semiconductors power key military and civilian capabilities relevant to national defense, such as certain military hardware and cloud computing.4 Moreover, advanced chips form the bedrock of recent advances in generative artificial intelligence (AI). For example, ChatGPT was trained on chips including NVIDIA’s 4nm H100, which as far as is known can only be manufactured in Taiwan.5
Secondly, the leading-edge node technology required to produce advanced chips is overwhelmingly located in Taiwan. Almost none of this technology is located in the United States.6 It is true that disruptions to non-advanced semiconductor supply chains pose a serious national security threat that ought to be examined further. However, advanced semiconductors should be the leading edge for American efforts to secure the supply chain because they underpin critical American military capabilities, form the bedrock of recent advances in AI, and represent where the United States is most dependent on foreign manufacturing.
Balancing Too Many Priorities
Current U.S. rhetoric around semiconductors advocates wielding vast reserves of American wealth to reshape the international supply chain in favor of American manufacturers.7 Many key government actors showcase this framing in their public statements, including Raimondo’s September 2023 comments above, the U.S. Department of Defense’s FY2021 Industrial Capabilities Report,8 and the White House’s press releases on the CHIPS Act, which explicitly state the goal of “keep[ing] America at the forefront of innovation and technological development”9 as well as “strengthen[ing] American manufacturing.”10
The presumptive end goal of such rhetoric—intentionally or not—is to replicate as much of the advanced semiconductor manufacturing ecosystem on U.S. soil as possible. Importantly, this goal is packaged as part of a wider program of industrial policy aimed at objectives such as boosting American competitiveness and creating domestic jobs.
This comprehensive U.S. approach is unwise for two reasons. First, it is likely to seriously harm the American relationship with Taiwan and put Taiwan at further risk. It is clear that a key part of Taiwanese strategic thinking involves its so-called “silicon shield.”11 Taipei sees their chokehold over crucial sectors of the semiconductor value chain as a pathway to embed Taiwan firmly within an economically interconnected web that keeps the island safe. The Chinese Communist Party (CCP) may perceive comprehensive onshoring as separating the interests of the United States and Taiwan, and therefore eroding Washington’s—not to mention other actors’—interests in defending the island.
Second, the economic inefficiencies involved with onshoring and the resulting distortion of the current market structure are likely to reduce the benefits of semiconductor supply chains to America. The United States’ position as the dominant end user of chips and its dominant position over IP and design software mean that it benefits the most from efficiencies in the value chain. Cheaper chips help American companies the most, especially when it comes to winning the R&D race in crucial industries such as national security technology and AI.
The semiconductor value chain benefits from massive efficiencies on the order of $100 billion12 from geographic specialization, which would be directly threatened by U.S. intervention to onshore and therefore de-specialize parts of the value chain. Moreover, distorting market incentives by unnecessarily creating greater uncertainty around costs is likely to further complicate decision-making by firms to invest. The Semiconductor Industry Association estimates that chips are between thirty-five to sixty-five percent cheaper because of the way the market is structured.13 Losing out on these cost savings is not a choice that decision-makers should make lightly.
Importantly, the true costs of a U.S.-based fab can only be understood by cyclically adjusting the running costs over the course of around two capital cycles. The term “capital cycles” refers to the observed phenomenon that individual firm-level decisions lead to aggregate expansions and downturns in industries.14 The semiconductor industry, in particular, is highly cyclical.15 As demand falls below short-term supply at the bottom of cycles, the most efficient fabs and producers are the most likely to survive, and the least efficient will collapse.
The industry’s trend towards consolidation and the fabless-foundry model has been accelerated by the result of punishing downturns.16 Chip firms are only likely to invest if they believe a plant will last through downturns. Decision-makers should therefore not labor under the misapprehension that spending to establish resilient semiconductor supply is a one-time deal; instead, the U.S. government is likely to have to offer an incentive to make up for the losses of fabs running below break-even utilization.
High Resistance
Besides being unwise, a comprehensive approach is unfeasible. There are at least three factors which will result in high resistance to a comprehensive approach. First, replicating semiconductor fabrication capacity onshore in the United States is likely to take too long. It has taken more than three decades to turn Taiwan into the kind of manufacturing hub which can conduct leading-edge R&D on a far lower cost basis than anywhere else in the world. Wholly relocating advanced chip manufacturing would not only require building expensive fabs in the United States but also relocation of the other crucial parts of the value chain. Building one advanced fab in Arizona is projected to take more than three years, so recreating the research ecosystem that enabled this technology would take far longer.17 It would certainly cost the federal government far more than the $50 billion currently apportioned by the CHIPS Act.
For perspective, Intel is currently building a singular high-volume outsourced semiconductor assembly and testing facility in Poland at a cost of $4.6 billion.18 Whereas a comprehensive approach is likely to require a timeline far too long to be effective for national security, a more concentrated approach that meets the two-to-four year timeline that is critical.
Second, it is exceedingly difficult to replicate the conditions which have led to the ability of Taiwanese companies to dominate the manufacture of advanced logic. It is at least thirty percent more expensive to run a fab in the United States than it is to run one in Asia.19 Higher costs in labor and electricity push up overall expenses, as well as increased annual depreciation costs because of the higher upfront capital expenditure required. By far the most significant factor is the difference in government policy across regions, which would require very comprehensive reforms to level. In an earnings call, TSMC said that the cost of construction of a fab in the United States could be at least four times the cost in Taiwan, driven in part by permits and regulatory compliance.20 There are non-subsidy reasons for the disparities in cost as well.
In Taiwan, the semiconductor industry enjoys entrenched talent pipelines. In 2021, the government set up four specialized semiconductor schools.21 To be sure, it is possible that such comprehensive reforms might pass in the United States, but it is difficult to imagine the enormous government expenditure required to put U.S. fabs on a level playing field with peers in East Asia. Even if the machinery of the U.S. government was so reformed, it is likely that higher operating costs would still weigh heavily on U.S.-based manufacturing because fabs are only one step in the value chain, albeit an important one.
Third, the United States occupies a dominant position in the end products of semiconductors, which means that it benefits most from efficiencies in the supply chain allowed by specialization, and loses out the most from any inefficiencies which are introduced. Consequently, in addition to being exceedingly difficult to restructure, the status quo should be maintained as much as possible, while still remaining compliant with crucial national security goals.
Taking a Concentrated Approach
The comprehensive approach rhetoric misses the mark. But the rhetoric is nonetheless animated by the correct impulse: that the United States should attempt to onshore a minimum viable capacity of advanced semiconductor production. A useful way to consider this minimum viable capacity is as an insurance policy that the U.S. government can take out against the security risks.
There are four main benefits to the concentrated, insurance-driven approach.
First, by insuring the United States against the threat of a Taiwan blockade or invasion, it spends no more than is absolutely necessary to protect America’s vital interests. Any dollar spent beyond this minimal insurance would be highly inefficient because of the aforementioned cost premium of moving production from East Asia to the United States—and, crucially, against America’s interests because it would raise the cost of the advanced semiconductors which are currently enabling AI development at U.S. companies.
Second, the insurance policy sends better signals to key actors. Taipei can keep as much of the silicon shield as possible, and Washington can affirm good faith with respect to its crucial Taiwanese partners.22 Specifically, concentrated policies allow Washington to signal that it is both potentially willing and able to defend Taiwan, in a more credible manner than either comprehensive policies or inaction will.
Third, given the smaller scope, the United States can address the national security concerns of advocates of comprehensive policies on a more realistic timescale.
Fourth, Washington is more likely to succeed in its efforts to achieve an insurance policy because it does not require replicating the entire complex ecosystem which makes Taiwan strategically important.
This kind of concentrated approach should be called an insurance policy because the model of insurance sheds light on the details of the capabilities which Washington should build. Any insurance policy has three questions associated with it: What is being insured against? What are the premiums? What payout is desired?
Insuring Against Adverse Events
Washington needs to safeguard itself against certain security risks, and thinking about semiconductor manufacturing capabilities through the lens of insurance will lead to smarter decisions about when the United States should pay premiums to insure against risk. There are a wide range of adverse events which could and should trigger the insurance to pay out. An invasion of Taiwan, attempted or completed, is at the forefront of many decision-makers’ minds. But even short of that extremity, it is feasible that a range of CCP actions would limit supply of chips from Taiwan and endanger U.S. access to the value chain.
Actions such as shipping interdiction, live-fire drills off the coast of Taiwan, or other CCP saber-rattling could, even if only temporarily, choke off chip supply to U.S. national security infrastructure. Finally, wise decision-makers should prepare for possible events that do not involve CCP action, but which still endanger semiconductor supply chains. One such case was 2020’s pandemic restrictions. Another is the possibility of another major earthquake in Taiwan which damages fabrication capabilities, such as in 1999 and 2016.
Calculating the Premiums
The high levels of capital expenditure required to stay competitive, let alone on the cutting edge, in the semiconductor market mean that any approach will carry large costs. Decision-makers therefore should limit these costs as much as possible because the United States benefits most from efficiencies in the supply chain—not to mention that the federal government is juggling priorities other than semiconductors. The upshot is that, akin to when one is buying insurance, the United States should keep premiums as low as possible while remaining commensurately insured against the risks to which it believes it is exposed.
Modeling an insurance model of this sort would enable decision-makers to adjust the parameters in order to determine a spending figure given a desired payout. Several key considerations would need to be addressed to build such a model. Specific considerations include the Department of Defense’s demand for advanced chips, the U.S. AI industry’s demand for advanced chips, the WFE conversion rate, fab monthly output, and the position of U.S.-based fabs on a global cost curve. On purely public information, these authors believe that offering a final cost would be insufficiently precise, if not misleading. Rather, we hope to offer an interactive version of our model that represents the implicit logic in an insurance policy. Decision-makers with access to this classified information could adjust the parameters of a model to answer these questions and design a concentrated approach intended to build an effective capability.
One final, crucial complication in modeling the premiums of an insurance model is that the nature of dual-use technology forces any approach to lie at the intersection of security and industrial policy. An appropriate approach must therefore incorporate industrial considerations. As noted above, in the case of a highly cyclical industry such as semiconductors, the insurance policy would need to offer expansive guarantees (subsidies or other support) through the capital cycle, softening the impact of industry downturns on dual-use fabs. The global nature of the industry and the long-term nature of investment is likely to lead to second-order effects throughout the supply chain. These effects will fall on allies and domestic firms alike, causing unquantifiable impacts outside the limits of our model. This sharpens the need for concentration in policy and clarity in communication with allies and industry.
Cashing in the Insurance
Decision-makers could reasonably pursue a wide range of outcomes with respect to semiconductor security. At a minimum, the insurance policy should guarantee the supply of advanced semiconductors for crucial national security priorities. With a contingency, any insurance policy should at least pay out in the ability to keep the U.S. Indo-Pacific Command fighting and the leading-edge chips that the Department of Defense needs in production.
Beyond this minimum, it should be up to decision-makers to decide how much they would like their insurance to pay out and therefore how high their premiums will be (in terms of the dollar figure that Congress will need to spend).
For instance, if decision-makers would like to retain the option to respond militarily to whatever event triggered the payout, then they should consider depletion of stocks in the event of a military contingency, which will necessitate correspondingly higher premiums. An approach which pays out to keep a comprehensive system viable through all downturns will have to pay incredibly high premiums to insure against those risks.
Becoming smarter about semiconductor supply means being precise about the amount that the federal government should spend to insure U.S. interests against specific risks. Instead of attempting a comprehensive approach which carries serious risks of not achieving national security goals, wise decision-makers should opt for concentrated policies which are clear-eyed about the risks that the United States could face in the coming decades.
References
1 The authors would like to thank Jason Hsu and Daniel Kroth for their assistance with the writing of this article. Alex Miller served as the lead writer, and Jake Birdwell oversees the research team between the University of Oxford and Stanford University which produced the analysis behind this article.
2 Michelle Toh, “TSMC Says Its $40 Billion Chip Project in Arizona Faces a Further Delay.” CNN Business, January 19, 2024.
3 Gina Raimondo, “Commerce Secretary Testifies on the Chips and Science Act.” CSPAN, September 19, 2023.
4 Sujai Shivakumar and Charles Wessner, “Semiconductors and National Defense: What Are the Stakes?” Center for Strategic & International Studies, June 8, 2022.
5 Hasan Chowdhury, “The hottest commodity in AI right now isn’t ChatGPT — it’s the $40,000 chip that has sparked a frenzied spending spree.” Business Insider, August 15, 2023.
6 Ibid.
7 “A 21st Century Defense Industrial Strategy for America.” U.S. Department of Defense, January 2021.
8 “FY2021 Industrial Capabilities Report to Congress.” OASD Industrial Base Policy, March 2023.
9 “FACT SHEET: One Year after the CHIPS and Science Act, Biden-Harris Administration Marks Historic Progress in Bringing Semiconductor Supply Chains Home, Supporting Innovation, and Protecting National Security.” The White House, August 9, 2023.
10 “FACT SHEET: CHIPS and Science Act Will Lower Costs, Create Jobs, Strengthen Supply Chains, and Counter China.” The White House, August 9, 2022.
11 Alexander Neill, “Doubts Grow Over Taiwan’s Silicon Shield.” GIS Reports, March 28, 2023.
12 “BCG x SIA Strengthening the Global Semiconductor Value Chain.” Boston Consulting Group & Semiconductor Industry Association, April 2021.
13 Ibid.
14 Hao Tan and John Mathews, “Cyclical Industrial Dynamics: The Case of the Global Semiconductor Industry.” Technological Forecasting and Social Change 77, no. 2, February 2010: 344.
15 Andre Rousselot, “Semiconductor market heads for biggest downturn since 2000.” eeNews Europe, September 13, 2022.
16 Ibid.
17 Cheng Ting-Fang and Lauly Li, “TSMC to send hundreds more workers to speed U.S. plant construction.” Nikkei Asia, June 29, 2023.
18 “Intel Plans Assembly and Test Facility in Poland.” Business Wire, June 16, 2023.
19 “BCG x SIA Strengthening the Global Semiconductor Value Chain.” Boston Consulting Group & Semiconductor Industry Association, April 2021.
20 Tim Mullaney, “How Manufacturing Chips in the U.S. Could Make Smartphones More Expensive.” CNBC, March 9, 2023.
21 Sarah Wu, “Taiwan Invests in Next Generation of Talent with Slew of Chip Schools.” Reuters, March 11, 2022.
22 Larry Diamond, Admiral James O. Ellis, and Orville Schell, Silicon Triangle: The United States, Taiwan, China, and Global Semiconductor Security. (Hoover Institution Press, 2023).