Thinking the Thinkable on an AI Market Correction
You’re not hallucinating: AI “bubble” discourse is everywhere. Whether you’re looking at Google Trends, reading the paper of record, listening to subject matter experts, or braving the Washington cocktail circuit, there’s a growing belief that AI will face some sort of market correction. Given AI’s strategic capabilities and the stakes of the technological competition with China, a potential market correction and its implications deserve scrutiny.While a “bubble” is not inevitable — or perhaps even likely — it is a real possibility. In the event of a market correction, powerful interests will intensify their calls for loosening key technology export controls and even offering concessions on Taiwan to Beijing. These measures would provide only ephemeral benefits at the cost of severely damaging long-term U.S. strategic and commercial interests. Bipartisan members of Congress must guard against short-sighted actions. The first priority should be to pass the Guaranteeing Access and Innovation for National Artificial Intelligence Act, which would prioritize American access to cutting-edge AI chips. Additionally, members of both parties should establish a statutory floor for export controls on China-bound AI chips, so that any substantial loosening requires explicit congressional approval. Finally, Congress should further institutionalize support for Taiwan. Uncertainties, Camps, and the J-CurveAnything is possible when it comes to AI development. Anyone who claims certainty is lying to you or deluding themselves. One useful way to think about the range of outcomes is to divide today’s AI debate into three schools of thought — sprinters, marathoners, and skeptics — and use that frame to assess how a market correction might unfold.Sprinters are the most optimistic of the three AI camps. In their view, AI capabilities (and financial market valuations) will continue their seemingly inexorable rise and potentially go exponential. Sprinters think a bubble is not imminent: In their view, the stock market is undervaluing AI-related stocks. Still, many sprinters have become less confident in recent months, with OpenAI chief executive Sam Altman edging away from prior artificial general intelligence projections. Other technology leaders have also grown more circumspect. While some AI companies may be hiding proprietary capabilities or even subtly encouraging bubble discourse (to muscle out weaker or less liquid competitors and consolidate the market), the sprinter camp seems much less self-assured than before.The skeptic camp, meanwhile, is having a moment. Skeptics like Gary Marcus point to OpenAI’s $1.4 trillion in spending commitments but only $13 billion (some say it’ll be closer to $20 billion) in annual revenue, or cite a Massachusetts Institute of Technology study finding that 95 percent of organizations have received zero measurable return from their generative AI projects so far. Mainstream outlets are increasingly reporting about a potential correction and identifying some AI companies increasingly engaging in creative financial engineering reminiscent of the Great Recession and circular investments. Many political figures from across the political spectrum, such as Florida Governor Ron DeSantis, are not necessarily skeptics but seem to be positioning themselves for the 2028 presidential election by expressing selective opposition to AI. Perhaps most concerningly, the bond market is increasingly showing wariness about repayment prospects.With U.S. total federal public debt-to-GDP ratios at roughly double pre-financial crisis levels, a correction could trigger significant near-term economic and financial pain even if AI holds substantial long-term value. Gita Gopinath, former chief economist at the International Monetary Fund, estimates that a correction would be disproportionately concentrated in the United States and eliminate $20 trillion in wealth for American households.Marathoners, meanwhile, acknowledge that a near-term market correction — perhaps even a deeply painful one — may be in the offing, but hold that AI will deliver significant or even transformative long-term benefits. As evidence, marathoners can point to studies from the St. Louis Federal Reserve finding significant productivity savings from generative AI or note that costs for inference (that is, model application) for a model at ChatGPT-3.5 capabilities fell 280-fold between November 2022 and October 2024.Marathoners hold that AI model capabilities could very well continue to “plateau” — delivering fewer marginal performance gains despite huge investment of incremental compute — but nevertheless improve productivity on a sector-by-sector basis. This scenario, similar to the “AI as normal technology” construct favored by AI experts Arvind Narayanan and Sayash Kapoor, would likely see long-term productivity gains from AI but with potential short-term growing pains.Indeed, Jason Furman, the former chair of the Council of Economic Advisers, warns that a “bubble” might be an imprecise way to describe a correction. Instead, Furman holds that an AI correction might resemble a “J-curve,” where AI adoption reduces productivity in the near-term but increases long-term output. Of course, if the downward slope of the J-curve proves steep, the consequences could be deeply painful.AI is already a powerful technology and holds transformative long-term potential. The United States should prepare for a long-term competition with Beijing and avoid making technological or geopolitical concessions to China even in the event of a painful J-curve market correction.The Aftermath of a Market Correction The shape of a potential market correction — whether a bubble, J-curve, or something else entirely — is difficult to predict but may strengthen Beijing’s leverage vis-à-vis Washington. In the event of a recession, U.S. policymakers will likely feel intense political pressures from higher unemployment levels and potential financial market contagion. Attention will turn to bolstering economic and financial linkages with China, the world’s second-largest economy. In such a scenario, Beijing will likely seek fewer export restrictions on advanced semiconductors and technological products.If a chip glut emerges, calls to soften export controls and “make a deal” will only grow louder. Analysts are split on a potential glut under current trends: some hold that oversupply is likely, while others are much more bullish. If a market correction significantly impacts demand, however, a chip glut becomes highly likely and would pressure manufacturers’ earnings. Indeed, in August 2025, and in an unprecedented arrangement that raised legal eyebrows, the Trump administration approved export licenses for Nvidia’s H20 and AMD’s MI308 inference-focused chips on the condition that the companies pay 15 percent of their China revenues to the U.S. government. Chip makers continue to openly advocate for softer restrictions — one hosted a major conference in Washington timed to immediately precede the Chinese General Secretary Xi Jinping-U.S. President Donald Trump Asia-Pacific Economic Cooperation meeting.In the event of a market correction, many in the U.S. business community will argue for further relaxing controls on inference chips, contending that selling less-advanced semiconductors will support an important domestic industry, pose little risk for America’s innovative edge, or even foster dependencies on the U.S. technology stack. Sales of inference or (especially) training chips to China will damage U.S. national security interests, however, perhaps severely.While inference chips are admittedly less sensitive than semiconductors used for training AI models, sales to China still pose major risks. Even less-advanced inference chips, such as the H20, could be repurposed to fine-tune training of powerful AI models. Figures in the administration have also floated potential sales of the significantly more powerful H200 chip, which could be used for either inference or training. As Chris McGuire of the Council on Foreign Relations notes, the H200 chip has over nine times more performance than current U.S. export control thresholds. Crucially, inference-capable chips would bolster China’s growing military edge in “physical artificial intelligence” across robotics and unmanned systems like quadrupeds or extra-large uncrewed underwater vessels. Selling inference chips to America’s most formidable military rival, even in the event of an economic crisis, is still deeply unwise.Selling more advanced chips used for training models would be even more dangerous, but cannot be ruled out. In the run-up to his autumn summit with Xi, Trump publicly floated allowing downgraded versions of Nvidia’s cutting-edge Blackwell training chips to be exported to China, before ultimately backing away under opposition from advisers and congressional China hawks.Industry holds that such sales would foster Chinese dependency on the U.S. technology stack, but this seems highly unlikely given Beijing’s track record, authoritative statements, and ambitions. In other technologies and industries — automobiles, batteries, solar panels, and even critical minerals — Beijing has run its industrial policy playbook of securing technology transfer, using scale to lower costs, and driving foreign competitors out of business. Moreover, AI and technology are central to Beijing’s Made in China 2025 plan and its 2017 New Generation Artificial Intelligence Development Plan — which explicitly aims to make China the global AI leader by 2030. Xi contended in a recent commentary that “Artificial intelligence is a strategic technology that is leading a new round of technological revolution and industrial transformation.” The Chinese leadership identifies AI as a strategic technology, in their own words. Consequently, they will seek to replace the U.S. technology stack rather than maintain dependency on it.Regardless of whatever it indicates publicly, Beijing seeks to acquire advanced chips currently subject to U.S. export controls. Of the main cost drivers over the model development process (research staff, accelerator chips, other server components, cluster-level interconnects, and energy), AI accelerator chips are China’s key bottleneck.If China can obtain enough advanced chips, however, its approach to AI development may change. Chinese AI industrial policy has emphasized sectoral adoption over developing bleeding-edge models. While there are several reasons for this approach, it has partly been necessitated by chip constraints. If U.S. chip controls are relaxed, then Beijing would be increasingly capable of sprinting to artificial general intelligence and might attempt to do so.Even if Beijing does not go all-in on developing artificial general intelligence, or if model capabilities plateau around current levels, China would likely use advanced training chips to develop the most powerful AI model that would outperform any American competitor. Chinese intelligence services would gain another useful tool for collection and influence, especially since Article 7 of China’s 2017 National Intelligence Law requires all organizations and citizens to “support, assist, and cooperate with national intelligence efforts.”If AI is truly a transformative technology, however, and artificial general intelligence is attainable, then China’s sectoral domination would all but ensure it would become the world’s preeminent and permanent power. Given the Communist Party’s devil-may-care safety record across COVID-19, the 2003 Severe Acute Respiratory Syndrome outbreak, and more, the consequences of its AI leadership could be profound.Taiwan as a “Bargaining Chip”?An AI correction will challenge Taiwan, as U.S. foreign policy is increasingly transactional and unmoored from traditional bipartisan calculations of long-term national interests. An AI market correction and resulting chip glut will simultaneously reduce Taipei’s leverage while increasing Beijing’s.An American AI recession — or worse — would not only refocus attention on U.S. ties with China, the world’s second-largest economy. It would also reduce Taiwan’s leverage by degrading the strategic and commercial value of its chip exports.If a glut of chips emerges, Taiwan will be hit hard. The semiconductor industry accounted for 13–15 percent of Taiwan’s GDP in 2023 (before the boom). About 317,000 individuals directly work in Taiwan’s semiconductor industry, or about 3 percent of its labor force. An AI market correction would therefore hold significant implications for Taiwan’s domestic political economy.While any recession would be wrenching, it is not the greatest danger facing Taipei. If the U.S. economy falls into a recession, would Taiwan be used as a bargaining chip? This is not an idle fear: U.S. policymaking is increasingly transactional, some U.S. policymakers hold little sentimental attachment to Taiwan’s hard-won democracy, and the Chinese economy is more than 20 times the size of Taiwan’s. For these reasons, an AI market correction may hold unprecedented dangers for Taiwan for two reasons. If the United States loosens chip controls, then China may develop global technological leadership. On the other hand, Washington could also conceivably and shortsightedly respond to a recession by using Taiwan as a bargaining chip with Beijing.To avoid either outcome, Taiwan and its friends should underscore that it is valuable to the United States for more than just chips. It is a strategically placed island and an advanced economy with deep technological expertise. Even leaving aside any sentimental notions of shared values, Taiwan’s absorption by China would alter the balance of power across the Indo-Pacific in Beijing’s favor. Many U.S. allies would hedge in the wake of a retreating leader and an emerging hegemon, leaving the United States increasingly alone to deal with a powerful rival that could summon all the resources and technological strength of mainland China and Taiwan.What Taipei Can Do NowAny camp — sprinters, marathoners, skeptics — could ultimately hold the correct long-term approach to AI. Still, policymakers on both sides of the Pacific should consider how they would respond to the very real possibility of a near-term market correction. Additionally, Taiwanese policymakers will have to navigate a complex U.S. policymaking process, cross-strait dangers, and their own fractious domestic politics.Taiwan can fundamentally bolster its resiliency by taking certain steps regardless of any AI market correction. Taiwan’s Public Debt Act, which caps the central government debt-to-GDP ratio at 40.6 percent, is inappropriate given manifest defense and energy security challenges. Taiwan’s domestic politics are admittedly highly polarized, constraining its ability to pursue needed reforms, but elites should understand they will fare much better in a constitutional democracy than under the Chinese Communist Party.Taiwan has already taken key steps to strengthen its resiliency, but more is needed. While defense spending will rise from 1.9 percent of GDP in 2021 to 5 percent of GDP by 2030, additional strategic investments in a “large number of small things” are necessary.Similarly, Taiwan can strengthen its energy security by pursuing an all-of-the-above approach to energy, using nuclear, liquefied natural gas, renewables, oil for generation and transportation, and even coal where appropriate.Getting Ahead of a Market CorrectionIn the event of a market correction, the U.S. response would prove critical. Powerful forces and interest groups will seek to relax U.S. export controls, weaken support for Taiwan, or both. Adopting either measure could cede technological leadership and geopolitical supremacy to the Chinese Communist Party.There are bipartisan measures that Congress can take now to reduce national security dangers in an AI market correction. Passing the bipartisan Guaranteeing Access and Innovation for National Artificial Intelligence Act will ensure that U.S. companies are first in line for advanced semiconductors. While export controls enjoy bipartisan support, codifying them via legislation of a statutory floor could greatly enhance their durability. Furthermore, substantially expanding the budget of the Bureau of Industry and Security would slow Chinese military and technological progress. Additionally, individual legislators could further signal bipartisan support for Taiwan by signing on to the already-large Senate and House Taiwan Caucuses. Legislators should prefer substantive support for Taiwan over showy displays and emphasize the one China policy, but they should also not shy away from upsetting Beijing.Resisting calls for loosening export controls or Taiwan-related concessions will be an all-hands-on-deck moment for the U.S. national security community. But if the United States and its allies and friends can overcome whatever complex dangers emerge from an AI market correction, a better world may emerge. Joseph Webster is a senior fellow at the Atlantic Council’s Global Energy Center and Indo-Pacific Security Initiative, and editor of the independent China-Russia Report. This article reflects his own personal opinion.Image: Gemini
