American AI Companies Can’t Get Enough Chips

This bottleneck has several implications for U.S. AI policy, with scarcity increasingly making AI chips a strategic resource. Every chip sent to competitors such as China is one less chip available to American AI companies, raising prices and slowing America’s AI progress. This is even the case for older AI chips, like NVIDIA’s H200, as it uses the same limited manufacturing capacity needed for more advanced AI chips.⁵ It also means one less chip available for the America AI Exports Program, which depends on sufficient supply to cement the U.S. tech stack in strategic third markets. Finally, it underscores the importance of frameworks like Pax Silica to not only tackle the upstream challenges of critical minerals but also build a more resilient, long-term allied supply of AI chips.

American AI Companies Can’t Get Enough Chips

The compute used to build and run AI systems has been growing exponentially, driven by continued growth in compute used for training and enhancing new models, conducting research and development (R&D) to improve the next generation, and deploying models at scale.⁶ Although AI models are becoming more compute efficient, this has not reduced aggregate demand for compute. Surging demand from AI scaling and adoption has overwhelmed efficiency gains.⁷

Compute constraints force AI companies to make difficult tradeoffs between tightening usage limits on customers, raising prices to suppress demand, cutting R&D investment, training smaller models, and serving a lower-quality product. Each option means less revenue, slower capability progress, and a weakened competitive position. These tradeoffs are already playing out. Recently, Anthropic introduced stricter rate limits on Claude during peak hours to manage demand.⁸ Google CEO Sundar Pichai claimed Google is “supply constrained even as we’ve been ramping up our capacity.”⁹ The leaders of major AI companies and their suppliers have consistently echoed this assessment (see the appendix).

These compute constraints are evident in AI chip rental prices. Historically, the cost of computing power has dropped exponentially over time, as more efficient hardware enters the market.¹⁰ However, analyses by Silicon Data and SemiAnalysis indicate that the rental price of the H100, a 2023 NVIDIA chip, is higher today than it was several years ago.¹¹ In other words, chip demand has more than offset efficiency-related price reductions.

Source: Center for a New American Security (CNAS), using data from Silicon Data and SemiAnalysis¹²

The tightest constraint for a given company will depend on its existing contracts and assets. However, as companies forward plan to scale up compute, their leaders consistently point to a specific bottleneck limiting the pace of compute buildout: manufacturing capacity for AI chips. According to an executive at Broadcom, which designs Google’s AI chips, “We are seeing that TSMC is hitting [production-capacity] limits. . . . They will be increasing the capacity to 2027, but that has become a bottleneck.”¹³ Taiwan Semiconductor Manufacturing Company’s (TSMC’s) CEO, C. C. Wei, echoed this claim: “The bottleneck is TSMC’s wafer supply, not the power consumption.”¹⁴

What Drives Chip Shortages

Chip manufacturers are reluctant to aggressively expand production for several reasons. First, long lead times, high capital costs, and boom-and-bust cycles are intrinsic to the business. Additionally, though, the memory of getting burned by overbuilding in response to inflated demand following the COVID-19 pandemic is still fresh. Automotive and consumer electronics manufacturers canceled orders early in the pandemic, anticipating a downturn. When the economy rebounded, they couldn’t secure adequate supply, as it had already been reallocated.¹⁵ This led to cars worth tens of thousands of dollars sitting unfinished on factory lots for want of chips costing only a few dollars, costing the automotive industry an estimated $210 billion in 2021.¹⁶ When manufacturers expanded capacity in response, they overshot actual demand as customers had been double-booking orders to ensure they got enough supply, leaving manufacturers with excess capacity and significant losses.¹⁷

The computer memory industry is especially susceptible to boom-and-bust cycles. Demand for memory can shift quickly, but new manufacturing capacity takes billions of dollars and many years to bring online. This timing mismatch can produce dramatic swings: High demand drives up prices and incentivizes investment, but by the time new capacity comes online, the industry has often overbuilt. These cycles have bankrupted company after company, consolidating the market from over 20 significant producers in the 1990s to just three main companies today: Samsung, SK Hynix, and Micron.¹⁸ These companies have seen demand spikes before and are wary of repeating mistakes that killed their competitors. These concerns make them hesitant to invest aggressively in response to surging demand from AI companies.¹⁹

TSMC, the Taiwan-based manufacturer that fabricates roughly 90 percent of the world’s most advanced chips, faces similar market dynamics.²⁰ The company’s CEO has acknowledged this directly: “You essentially try to ask whether the AI demand is real or not. I’m also very nervous about it. . . . If we did not do it carefully, that will be a big disaster to TSMC for sure. . . . I want to make sure that my customers’ demands are real.”²¹ To hedge against the risk of evaporating AI demand, TSMC is also reserving capacity for customers with a long track record of reliable demand, such as Apple, even if that means accepting lower prices.²²

This caution has directly contributed to the current AI chip bottleneck. Despite the surge in spending on AI chip investments after the release of ChatGPT, capital expenditures from TSMC and major memory manufacturers were lower in 2023 and 2024 than in 2022 (Figure 2). While they are investing more aggressively now, it takes several years to build new manufacturing capacity. The concentration in the semiconductor industry means that there are not viable alternatives if demand spikes. Therefore, AI chip production is effectively capped by how much a few, cautious companies have invested in the past few years.²³

Source: CNAS, using data from public financial filings, Epoch AI, and forecasts for 2026 CapEx²⁴

AI-driven demand, meanwhile, has continued to skyrocket. As of April 2026, Anthropic’s annualized revenue has surged to $30 billion, up from $9 billion just four months earlier.²⁵ This has caught chip manufacturing companies flat-footed. Reportedly, NVIDIA and Broadcom requested additional manufacturing capacity from TSMC, only to be turned down. Google has reportedly been unable to increase its AI chip production to meet its 2026 targets because it did not secure enough manufacturing capacity.²⁶ These chip manufacturers are now investing more aggressively in additional capacity, but given the time required to bring new capacity online, the availability of AI chips will likely continue to be the tightest constraint on AI scaling in the near term.²⁷

TSMC’s first Arizona fabrication facility (fab) is now producing four-nanometer (nm) chips, with five more fabs planned through a $165 billion investment.²⁸ As the demand signals from AI companies work their way through the supply chain, other companies are making bets to increase chip supply, including xAI CEO Elon Musk’s ambitious goal of building his own fabs.²⁹ Market forces will eventually boost supply to meet the demand, but that will take years to bear fruit. Tightness in the chip supply chain will remain through at least the end of 2026.

Specific Bottlenecks Within AI Chip Production

AI chips consist of several distinct components that share many manufacturing components with each other and with consumer hardware, such as smartphones. Key components include logic dies (which perform the computations) and memory, which is packaged with the dies on the same AI chip.

Figure 3 | Memory and Logic Wafer Production for AI Chips Is Highly Concentrated

The production of logic dies and memory is particularly tight. Both rely on semiconductor fabs with highly sophisticated “clean rooms” to manufacture chips with ultraprecise lithography machines. These environments must be kept more pristine than a hospital operating room, as a single speck of dust can destroy chips during production.³⁰ Building new clean rooms takes years, which is a main reason that production of logic and memory cannot respond quickly to surges in demand.³¹

Logic Wafers

Finished logic wafers—the processed silicon discs from which individual logic dies are cut—are currently one of the tightest constraints on AI chip supply. Logic dies provide the core processing component of an AI chip, but logic wafers are also used to make networking chips, central processing units (CPUs), and parts of memory chips.

NVIDIA, AMD, and other AI chip designers rely on TSMC’s world-leading, advanced node processes to fabricate their logic chips. But TSMC’s manufacturing capacity is finite in the short term and increasingly oversubscribed, given the surging demand for AI compute. As a result, even TSMC’s largest customers are short of supply.³²

In November 2025, TSMC’s CEO stated that the company’s advanced process capacity was “not enough, not enough, still not enough” and that demand was running roughly three times ahead of what the company could produce.³³ He also reportedly joked about wearing a shirt that read “no more wafers” to emphasize the severity of the shortage.³⁴ TSMC’s production capacity for 3 nm chips stood at around 70 percent utilization in early 2025, but it has been near and even above 100 percent since late 2025, as the company pushes machines above their planned capacity, including by delaying maintenance.³⁵ The company’s 3 nm node is especially constrained because it produces today’s most advanced AI chips, including NVIDIA’s Vera Rubin and Google’s TPUv7.³⁶ The company’s 2 nm fabrication capacity is booked through 2028.³⁷ TSMC cannot meaningfully ramp up advanced wafer manufacturing capacity in the near term because constructing additional fabs takes two to four years and sometimes longer.³⁸ New capacity currently under construction is also unlikely to resolve the shortage. TSMC’s planned Arizona Fab 4 is already fully booked and the ground hasn’t even been broken yet.³⁹

Source: CNAS adaptation from SemiAnalysis⁴⁰

Other industries, such as smartphones, have historically consumed significantly more advanced logic wafer capacity than AI chips, so AI companies increased their chip production by simply outbidding those industries for production slots. However, with AI compute production growing at over three times per year, it is beginning to run up against TSMC’s total manufacturing capacity for certain manufacturing lines.⁴¹ Flagship customers such as Apple and Broadcom have explicitly stated that they are constrained by the available supply of TSMC chips, while the CEO of TSMC stated on the April 2026 earnings call that it will not be until 2027 that “supply can meet demand.”⁴²

Memory

Memory is another bottleneck. Smartphones and personal computers use dynamic random-access memory (DRAM) as their standard form of working memory. AI data centers use both this conventional DRAM and a specialized variant called high-bandwidth memory (HBM), which stacks multiple DRAM dies vertically to deliver the massive bandwidth AI chips require. When AI models are deployed, performance is shaped by how quickly the chip can move data through its memory, which is the key functionality HBM provides. As more computing power is devoted to running models at scale, HBM has grown increasingly important and scarce.⁴³

For a time, this wasn’t a problem. DRAM is a large market that predated the AI boom, so AI chipmakers could get more memory by simply buying a larger share of existing production. However, similarly to the case with advanced logic wafers, that runway has narrowed as AI memory demand has skyrocketed. Aggregate HBM bandwidth has increased by over four times per year, and AI now accounts for the majority of DRAM demand (Figure 5).⁴⁴

Source: Adapted from SemiAnalysis⁴⁵

Growing demand for HBM has stressed the production capacity of the entire memory industry, causing some DRAM prices to increase over 600 percent in 2025.⁴⁶ In addition to DRAM (used for working memory), demand for long-term memory for AI servers drove up the price of NAND Flash (the memory used for long-term storage) over 300 percent in 2025.⁴⁷ Because DRAM and NAND are also essential for consumer electronics like smartphones and laptops, these industries have been hit hard by the price shock.⁴⁸ Industry analysts expect the memory shortages and price increases to shrink the global PC and smartphone markets by 11 and 13 percent, respectively, in 2026. In response to the tightness in DRAM memory supply, Meta recently stated that they would extend the planned lifetime of some of their older data centers, stating: “New server procurement cannot keep pace with demand growth in the near term.”⁴⁹

While certain buyers can get more memory by outbidding others, overall memory production is finite in the short term. The AI compute buildout cannot grow faster than memory fabrication capacity allows. For both memory and logic wafers, production capacity is also often locked into longer-term contracts, especially relative to the pace of the AI industry, meaning that additional capacity is unavailable at any price.

Several factors compound the current memory shortage. First, making HBM requires three to four times as many memory wafers per gigabyte of memory as standard DRAM does.⁵⁰ When chipmakers expand HBM production, therefore, it comes at a steep cost to conventional DRAM supply. Second, the amount of HBM packed into each AI chip is rising exponentially with each generation.⁵¹

Third, memory manufacturers cut investments in new fabrication facilities in 2023 and 2024, when they were losing money on memory (Figure 2).⁵² They are increasing investment in 2026, but it will take years for new memory fabs to come online. Finally, improvements in memory density have stalled over the past decade, meaning the industry can no longer count on the efficiency improvements that historically allowed it to dramatically grow output without building as many new fabs.⁵³

These factors have combined to cause what an executive at Micron has described as “the most significant disconnect between demand and supply in terms of magnitude as well as time horizon that we’ve experienced in my 25 years in the industry.”⁵⁴ Despite investments in additional manufacturing capacity, Micron CEO Sanjay Mehrotra has stated that “aggregate industry supply will remain substantially short of the demand for the foreseeable future,” and SK Hynix CEO Kwak Noh-jung has stated that the “current shortage could continue until 2030.”⁵⁵

Demand for AI chips is straining manufacturing capacity for both advanced logic and memory, limiting the rate of America’s AI buildout. Given the inelastic nature of these supply chains in the short term, these constraints will persist through 2026.

Other Tightness in the Chip Supply Chain

Other constraints to the AI buildout could soon emerge as the technology continues to scale and evolve. For example, as AI handles more agentic tasks that require CPUs rather than graphics processing units (GPUs), such as web browsing, shortages of CPUs are likely as the market adapts.⁵⁶

Advanced packaging, the process of integrating compute dies and HBM stacks onto a single substrate, was a binding constraint on AI chip production in 2023.⁵⁷ The dominant advanced packaging technology is TSMC’s Chip-on-Wafer-on-Substrate (CoWoS), which is used in nearly every major AI chip.⁵⁸ Since 2023, CoWoS capacity has expanded significantly, with SemiAnalysis characterizing it as “tight but easing” as constraints on memory and logic wafers have become relatively more severe.⁵⁹ With that said, packaging constraints have yet to fully disappear. CoWoS capacity was reportedly the bottleneck that forced industry analysts to revise their forecasts of Google’s 2026 AI chip production plans from four million to three million chips.⁶⁰

At the company level, the most binding constraint is a function of existing contracts, supply agreements, and capital planning. For example, in late 2025, Microsoft continued to highlight energy as its primary concern, while in early 2026 OpenAI executives stated that their binding constraint had shifted from power to chips.⁶¹ However, these constraints also interact dynamically. If a company expects to be power constrained, it will not invest as aggressively in securing chips that it cannot run, which in turn weakens demand signals for chip manufacturers to expand capacity. But some industry analysts expect that, in aggregate, chip constraints, rather than power constraints, will be the binding constraint on AI in 2026 and beyond.⁶²

Taken together, these supply chain constraints make AI chip production a significant bottleneck on AI progress for 2026 and potentially for years to come.

Policy Implications

Given the constraints on AI chip supply, the United States has both greater leverage and greater reason to ensure every chip is put to its highest-value use.

Rising Costs May Threaten Beneficial Uses of Compute

Increased demand and constrained supply for AI chips threaten to keep AI compute prices elevated for the near future. Well-funded AI companies can likely absorb the premium, but researchers and academics risk being priced out of access to the computing power needed for foundational research and innovation. To mitigate this risk, the U.S. government should increase funding for compute subsidies through the National AI Research Resource initiative (NAIRR). NAIRR seeks to provide American researchers with access to computing power and other resources for AI research and AI-enabled scientific discovery. Congress has approved $30 million to continue the program through fiscal year 2026—a year in which hyperscalers are projected to spend around $690 billion in AI capital investments.⁶³ This year, for every dollar allocated to NAIRR, the private sector is investing roughly $23,000 in AI.⁶⁴ To offset rising compute costs and ensure access for U.S. researchers, students, and small businesses, Congress should increase NAIRR’s funding and ensure AI-enabled science and innovation continue apace.

Exporting Chips to Autocracies Such as China Undercuts U.S. and Allied Chip Access

The bottlenecks connected to AI chip supply means that current chip allocation is zero-sum. Every chip exported to an authoritarian regime is one less chip available to firms in the United States and allied democracies to train a better model, serve more customers, or strengthen the democratic AI ecosystem. In a supply-constrained environment, the United States and its allies should treat chip access as a strategic resource and prioritize exports to democracies over authoritarian states that are more likely to deploy AI in ways that undermine U.S. interests and democratic norms. Without location verification on chips or other monitoring mechanisms, there is no way to confirm that exported chips are not being diverted to harmful actors within these nations. Even if exported chips do not directly reach harmful actors, increasing overall chip supply in a severely chip-constrained country such as China frees up marginal compute that can be redirected to these actors.

The direct tradeoffs of chip exports to China are clear. In March 2026, when regulatory uncertainty stalled H200 sales to China, NVIDIA redirected TSMC capacity from H200 production to its next-generation Vera Rubin chips, which had confirmed orders from OpenAI, Google, and other American firms.⁶⁵ As one person familiar with the decision put it: “Nvidia has to move on to what it can achieve with certainty, especially when there’s a shortage of supply for its advanced stuff.”⁶⁶ Less-advanced AI chips like NVIDIA’s H200 consume the same constrained manufacturing capacity as current frontier AI chips (including TSMC wafer capacity, HBM, and advanced packaging), meaning they too come at a direct cost to U.S. and allied supply.⁶⁷ The Bureau of Industry and Security rule permitting the sale of H200 chips to China requires exporters to certify that shipments will not divert capacity from the U.S. market. But when AI chips all draw on the same finite pool of upstream capacity, it is difficult to see how any company could credibly certify that chip exports to China do not come at the expense of American supply.⁶⁸

Some stakeholders argue that allowing limited chip sales to China can slow domestic Chinese chip development by keeping Chinese companies dependent on American hardware. However, Beijing is already pouring billions into domestic chip production, and chips alone do not create lasting lock-in.⁶⁹ While China is currently constrained in the quantity and quality of AI chips that it can produce, every U.S. chip sent to China supports the buildout of a competing AI ecosystem that will eventually challenge American firms in third-country markets.⁷⁰ Continued sales risk eroding the very advantage that the AI Exports Program and Pax Silica partnerships are designed to secure.

Exporting the American AI Stack Remains a Critical Imperative

Blocking AI chip sales to China does not mean the United States should hoard them from the world. Exporting AI chips to allies directly benefits U.S. companies and can reinforce America’s AI advancements.⁷¹ Most chip exports involve American firms building data centers abroad, which in turn allocate that compute however is most effective. For instance, they can use the compute to deploy AI services to consumers to generate revenue, advance internal R&D, or generate synthetic data for training new AI models. When AI industry executives and analysts identify chips as the tightest constraint, they are already factoring in data center buildouts abroad and in allied countries. Restricting exports to allies would threaten that capacity, potentially making the availability of powered data centers the binding constraint again on the AI buildout.

Additionally, exporting AI and its requisite computing power is vital to America’s AI leadership. Failing to export American AI while China ramps up capacity to eventually serve third markets risks repeating the mistake of 5G, where the United States pioneered the underlying technology but ceded global deployment to Huawei.⁷² This error took years and billions of dollars to begin unwinding. The United States should export strategically in line with the AI Exports Program, locking in an ecosystem advantage over Chinese competitors and deepening partnerships with allies to secure long-term leadership.

Smuggling Matters More When Chips Are Scarce

Chip shortages add urgency to countering chip smuggling. Every smuggled chip means fewer chips available for the United States and its allies, undermining America’s AI progress at home and ability to export the American tech stack to the world. Washington needs to step up enforcement, including through technical innovations that make tracking cheaper and more scalable. The bipartisan Chip Security Act, which passed the House Foreign Affairs Committee unanimously in March 2026, would require location verification mechanisms for exported AI chips and mandatory reporting of suspected diversions.⁷³ Passing this bill would make export controls and their enforcement more effective.

Washington should also consider updating its approach to export controls to better protect scarce resources, specifically by adopting a whitelisting approach for HBM exports. HBM is an unusually well-suited target for export controls because the number of legitimate buyers is extremely small.⁷⁴ Only a handful of companies worldwide need raw HBM: those designing GPUs, AI accelerators, and other high-performance chips. A whitelist approach, restricting HBM sales to approved chip designers, would ensure that HBM ends up integrated into chips that serve the U.S. and allied market rather than being diverted to competitors.

Coordinating with Allies Can Help Strengthen Supply

U.S. and allied governments are already subsidizing and shaping semiconductor supply chains.⁷⁵ But without joint planning, there’s a risk that every government will chase the same bottlenecks while underinvesting in less visible upstream inputs. This could result in wasteful duplication in some areas and persistent gaps in others. Pax Silica can serve as the coordination layer, aligning allied investment so that each country builds on existing strengths and collectively delivers an allied semiconductor base that is sufficient and not vulnerable to any single point of failure.

The United States and its allies hold a commanding lead in AI chips, but it will not last forever. Given the shortage of AI chips, the United States should export aggressively to countries that strengthen its AI leadership and restrict exports to those that would build a competing ecosystem. Washington has a narrowing window in which to shape the global AI infrastructure buildout, deepen allied partnerships that reinforce long-term U.S. leadership, and anchor AI development in democratic values. Leveraging that window requires government and industry to work together and treat scarce chips as the strategic resource they are.

About the Authors

James Sanders is a research associate for the Technology and National Security Program at CNAS. His research focuses on the implications of AI, including how compute policy can be used to manage the risks and opportunities of advanced AI systems. Before CNAS, Sanders worked for Epoch AI studying trends in AI capabilities and infrastructure and worked as a quant trader at Susquehanna International Group. He holds a BA in mathematics from Rice University.

Janet Egan is a senior fellow and deputy director of the Technology and National Security Program at CNAS. Her research focuses on the national security implications of AI, including how compute policy can be used to manage the risks and opportunities of advanced AI systems. Prior to joining CNAS, Egan was a director in the Australian Department of the Prime Minister and Cabinet. Egan holds a master’s in public policy from the Harvard Kennedy School and a BA from Monash University in Australia.

Rory Madigan is a 2026 spring research fellow with the Cambridge Boston Alignment Initiative. He was previously an associate at the private investment firm Crane Partners. Madigan holds a holds a master’s in global affairs from Tsinghua University in Beijing and a BA from Columbia University.

About the Technology and National Security Program

The Technology and National Security Program produces cutting-edge research and recommendations to help U.S. and allied policymakers responsibly win and manage the great power competition with China over critical and emerging technologies. The escalating U.S.–China competition in AI, biotechnologies, next-generation information and communications technologies and digital infrastructure, and quantum information sciences will have far-reaching implications for U.S. foreign policy and national and economic security.

The program focuses on high-impact technology areas with in-depth, evidence-based analysis to assess U.S. leadership vis-à-vis China, anticipate technology-related risks to security and democratic values, and outline bold but actionable steps for the United States and its allies to lead in responsible technology development, adoption, and governance. A key focus of the program is convening the technology and policy communities to bridge gaps, exchange perspectives, and together develop solutions.

Acknowledgments

The authors would like to acknowledge Georgia Adamson, Vivek Chilukuri, Liam Epstein, Tim Fist, Isabel Juniewicz, Michelle Nie, Paul Scharre, Josh You, and others who provided valuable feedback and insights throughout the development of this report. The views expressed in this paper do not necessarily represent those of acknowledged individuals nor affiliated organizations. The authors are also grateful for the CNAS publications and communications teams for their support and editing. This paper was made possible with the generous support of Coefficient Giving.

As a research and policy institution committed to the highest standards of organizational, intellectual, and personal integrity, CNAS maintains strict intellectual independence and sole editorial direction and control over its ideas, projects, publications, events, and other research activities. CNAS does not take institutional positions on policy issues and the content of CNAS publications reflects the views of their authors alone. In keeping with its mission and values, CNAS does not engage in lobbying activity and complies fully with all applicable federal, state, and local laws. CNAS will not engage in any representational activities or advocacy on behalf of any entities or interests and, to the extent that the Center accepts funding from non-U.S. sources, its activities will be limited to bona fide scholastic, academic, and research-related activities, consistent with applicable federal law. The Center publicly acknowledges on its website annually all donors who contribute.