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.

Anything 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.