US steps up AI investment, why is the capital market divided

On October 17, at the 2026 Annual Meetings of the International Monetary Fund (IMF) and the World Bank Group, Anthropic co-founder Clark said that artificial intelligence (AI) capabilities are improving rapidly, and according to its internal model estimates, U.S. GDP could grow by 50% by 2030, driven by AI. But despite such an attractive outlook, U.S. capital markets are showing two different sentiments: on one hand, they continue to increase industrial investment; on the other, they worry about overheating investment and a market bubble. Are investment returns overestimated? In recent years, U.S. AI investment has maintained strong growth momentum. Statistics show that the capital expenditure of five hyperscale cloud providers is expected to approach $800 billion this year. Behind this is the market's high expectations for AI's development prospects and its belief that AI is creating a huge new market. The bullish side believes that the current large-scale investment is a necessary move to seize future industrial advantages. U.S. asset management company BlackRock still lists AI as one of its most favored investment themes, believing that the computing facilities and R&D systems currently being built do indeed take a long time to pay back, but may become an important foundation for future economic growth. IMF Managing Director Georgieva said at a meeting held in New York on September 20 that although the world economy is more resilient than many people expected, it still faces risks from stubborn inflation, rising debt servicing costs and overheating AI investment. Xinhua News Agency file photo But the other side believes that this continuously increasing investment is borrowing from the future in advance. Ray Dalio, founder of U.S. investment firm Bridgewater, has called the AI boom a "classic bubble," arguing that large-scale borrowing combined with rising interest rates is approaching the conditions for a bubble burst. Bank of America's October global fund manager survey showed that nearly half of respondents believe a bubble will appear next year. Some companies have already borne high debt costs to seize the computing power market. Computing power leasing company CoreWeave's net interest expense in the second quarter of this year reached $640 million, equivalent to about a quarter of its revenue. Credit market forecast data show that CoreWeave's probability of default in the next five years is as high as 63%, and its stock price has also recently suffered a sharp setback. Although market judgments are clearly divided, fund managers' net overweight in tech stocks has still risen to the highest level since August 2020. This seemingly contradictory phenomenon reflects that investors are both worried that AI is being overpursued and unwilling to miss the growth dividends brought by this round of technological change. Neither of the two diametrically opposed judgments denies AI's technological prospects; rather, they focus on whether the large amount of current investment and high valuations can be supported by real returns. Understanding the U.S. AI boom still requires judgment from two dimensions: industrial supply and end-user applications. Overcapacity may emerge in the future. From the perspective of short-term supply, U.S. computing infrastructure remains tight, existing construction projects are supported by real demand, and a large amount of new capacity has already locked in customers before it is completed. A report by real estate services firm CBRE shows that in the first half of this year, supply in major North American data center markets increased by 33.7% year on year, but the vacancy rate was as low as 1.4%, and 80.4% of capacity under construction had already obtained pre-lease commitments. But the current computing power shortage and future overcapacity are not contradictory; both may appear in the same cycle. As a heavy-asset project, data center construction takes years and cannot turn on a dime as market demand changes. The Bank for International Settlements pointed out in its "Annual Economic Report 2026" that precisely because of current bottlenecks in electricity, advanced chips and power grid equipment, companies are rushing to lock in forward capacity through long-term contracts to avoid falling behind. Furthermore, the forward orders companies sign may become decoupled from real future demand. Some forward demand has already cooled. A report by the market monitoring agency of PJM, the largest regional power grid in the United States, said that about 42% of the previously forecast additional power load from data centers has been canceled or postponed. In addition, the large-scale expansion of U.S. computing infrastructure is also constrained by social costs and local policies. In February this year, hundreds of people gathered outside San Marcos City Hall in Texas to protest a proposed $1.5 billion data center. Residents expressed opposition at a hearing, mainly out of concern about the project's water consumption, electricity use and long-term environmental pressure. The city council subsequently voted down the land-use zoning change required for the project. This is not an isolated case. New York State also announced in July that it would suspend the issuance of state environmental permits for data center construction with power capacity of 50 megawatts or more. User perceptions return to rationality. From the end-market perspective, AI is being widely used in areas such as software development, office collaboration and customer service. More and more enterprises are willing to pay for AI productivity, and the revenue of leading model companies is also growing rapidly. Reports say OpenAI's annualized revenue this year is expected to exceed $50 billion, and its research on individual ChatGPT users shows that after six months of registration, users' average daily message volume is about 50% higher than in the first month. But as model applications penetrate traditional workflows, enterprises' attitudes are also becoming more rational. An EY survey of U.S. executives shows that 82% expressed concerns about token usage and related costs. A S&P Global survey shows that the proportion of enterprises "abandoning model projects midway" has doubled compared with the previous round of statistics. As model efficiency improves and market competition intensifies, model invocation prices continue to fall, and technological progress has instead narrowed profit margins to a certain extent. Well-known venture capital firm a16z estimates that driven by algorithm optimization and the spread of open-source models, the inference price of large models with equivalent performance is falling by about 90% per year on average, increasing pricing pressure on model service providers and also affecting profit margins at all links in the industrial chain. On July 23, 2025, the White House released an artificial intelligence (AI) action plan, listing dozens of federal policy actions the Trump administration will take in the future. Some media analyses believe that this move means the Trump administration is rescinding multiple policies from the Biden administration era and loosening regulatory measures on AI development. Xinhua News Agency released this. The American public's perception of AI risks is also shifting. The "Future Risks Report" released by AXA and Ipsos shows that among the risks ranked by surveyed experts for the next five to ten years, AI and big data risks surpassed climate change to take first place. In the latest tracking survey by the Pew Research Center, 52% of American adults feel "more concerned than excited" about the use of models in daily life, and as many as 71% of respondents expect AI to reduce jobs in the next two decades. A nationwide poll by Quinnipiac University further shows that 74% of Americans do not trust the leadership of model companies, 77% believe that the development of powerful model systems should be paused or slowed until safety is evaluated, and 86% support enforcing independent safety standards even if it slows down research and development. This widespread mindset in American society provides a foundation for groups that insist on "being bearish" and "slowing down." Judgment requires multidimensional observation. However, the Trump administration, which adheres to the logic of "racing," has always clearly suppressed all "slowdown" arguments. In the face of the recent "slowdown theory" called for by Anthropic CEO Dario Amodei and others, the Trump administration has given a diametrically opposite tough response, firmly anchoring the policy focus on increasing investment and winning the artificial intelligence race. Trump himself alleged that "so-called AI doomsday warnings are an exaggerated hoax." President's Science and Technology Advisor David Sacks, who once served as the White House AI "tsar," was even quicker to publicly mock leading companies for seeking "regulatory capture" and industry monopoly in the name of safety. This "racing" logic has driven a series of policy trajectories of the Trump administration since 2025. On one hand, the White House is using "Winning the Race: America's AI Action Plan" to clear regulatory barriers, accelerate approvals for data centers and electricity, and promote full-stack exports of American AI technology; on the other hand, it has launched the "National AI Legislative Framework," urging Congress to replace fragmented restrictive bills in various states with unified standards and accelerate the establishment of a nationally unified federal AI policy framework with the least burden, striving to lock in America's absolute dominance through comprehensive deregulation combined with accelerated infrastructure construction. Therefore, judging the American AI boom or bubble cannot rely solely on observing a single dimension, but requires deeper and longer-term observation of the tensions triggered by AI in the American market, society, and political system. (Yao Xu is Secretary-General of the Global Center for AI Innovation and Governance and Associate Researcher at the Fudan University Development Institute; Guo Haobo is a research assistant at the Global Center for AI Innovation and Governance.)