AI Bubble

Structure of the argument

The AI bubble thesis holds that the market values of AI-related firms have been inflated well beyond their actual capacity to generate revenue and the maturity of the technology, with the bubble concern growing since 2025. The argument rests on three grounds.

The first is the gap between investment and revenue: training and running large language models requires enormous GPU and power costs, while actual AI service revenue falls short of them, an 'AI profitability gap'. David Cahn of Sequoia Capital pointed this out in his June 2024 report 'AI's $600 billion question', and Sequoia judged the gap between AI infrastructure investment and revenue to be around $500 billion or more, unlikely to close in the short term. The second is circularity in investment. Vendor financing is the practice by which chipmakers and cloud providers provide capital, guarantees or purchase commitments to firms buying their hardware and services, making chipmakers, cloud operators and frontier labs one another's investors, suppliers and customers. The third is technical limitation. The cognitive scientist Gary Marcus criticised hallucination as a fatal obstacle to industrial adoption and argued that current investment is largely speculative rather than grounded in fundamentals.

Circular dealing takes the form of a chipmaker or cloud provider taking equity in, or extending finance to, a firm that has agreed to buy its products and services, with the multi-billion-dollar contracts of Oracle and CoreWeave cited as examples. Morgan Stanley estimates that debt used to finance data centres could exceed $1 trillion by 2028, and much of the related debt is rated BBB or junk. At the end of 2025, 30 per cent of the S&P 500 and 20 per cent of the MSCI World index rested on just five companies, the highest concentration in half a century, alongside a forward price-to-earnings ratio of 23 for the S&P 500 (against 14 for the FTSE) and a US Shiller PE above 40 for the first time since the dot-com collapse.

Historical background

The debate is set against two 'AI winters' in artificial intelligence research. The perceptron boom of the 1950s and 1960s ended when Minsky and Papert demonstrated the limits of early neural networks in 1969 and the 1973 Lighthill report cut off government funding. The expert-systems boom of the 1980s revealed brittleness, being hard to maintain, unable to learn new data and lacking commonsense reasoning, and collapsed in the late 1980s as desktop PCs improved and destroyed the dedicated-hardware market. The third boom, which began with deep learning in the 2010s, exploded when OpenAI released ChatGPT in November 2022 and prompted massive data-centre investment by Microsoft, Google, Meta and Amazon. The bubble debate that followed is posed as a question of whether this is a repetition of the cycle or a paradigm shift, and Gartner's 2024 hype cycle report judged that generative AI was passing the 'peak of inflated expectations' and entering the trough of disillusionment.

Specific events drove the debate in 2025. In January, the Chinese-built chatbot DeepSeek showed comparable performance at far lower cost; Nvidia fell 17 per cent in a day, losing about $600 billion in market value, then recovered 8.8 per cent the next day. In July, Nvidia became the first company to reach a $4 trillion market capitalisation, four times its 2023 level of $1 trillion and about 7.3 per cent of the S&P 500. In October it passed $5 trillion, and over 2025 AI-related firms accounted for roughly 80 per cent of the rise in US markets.

That year the circular structure appeared in concrete cases. In September Nvidia announced a $100 billion investment in OpenAI, forming a circuit in which OpenAI runs data centres on Nvidia GPUs. In October OpenAI bought billions of dollars' worth of GPUs and CPUs from Nvidia's competitor AMD, becoming one of AMD's largest shareholders, while Microsoft and Oracle (a $300 billion contract) were also entangled in the structure.

Institutional warnings followed. The Bank of England warned that overvaluation of major AI technology stocks was raising the risk of a global market correction, adding that if AI infrastructure costs are too high valuations could fall further and that investors had not been properly warned of the risk of a stock market crash. The International Monetary Fund supported the warning, and its managing director Kristalina Georgieva compared the situation to the dot-com bubble and said a correction could slow world growth and weaken developing economies.

The counter-argument

The opposing case is also substantial. Goldman Sachs pointed to the absence of a killer app and physical constraints such as the power grid while also presenting the positive projection that AI could raise US productivity by more than 9 per cent and GDP by 6.1 per cent over a decade, reaching mixed conclusions. Peter Oppenheimer of the same firm countered that the share price rises of large AI firms rest on real earnings growth and that forward price-to-earnings ratios are lower than in the dot-com era. Morgan Stanley analysts called bubble fears wrong or premature, noting that the median cash flow and capital reserves of the top 500 US companies are about three times higher than in past bubble periods. JP Morgan concluded that AI does not meet the classic criteria for a bubble, and Federal Reserve chair Jerome Powell likewise distinguished it from the dot-com era on the ground that AI firms generate real revenue.

At the same time, some voices concede the possibility of a bubble. Ray Dalio said in early 2025 that current AI investment levels are 'very similar' to the dot-com bubble, and Sam Altman said in August 2025 that he believed an AI bubble exists. Jamie Dimon of JP Morgan said in October 2025 that 'AI is real' but that some money invested will be wasted and there is a higher chance of a meaningful stock drop over the following two years than the market reflected. SEC chair Gary Gensler warned of so-called AI washing, the exaggeration of simple algorithms as 'AI-powered'. A February 2026 study by the National Bureau of Economic Research found that while 90 per cent of firms reported no effect of AI on their workplaces or productivity, executives projected productivity gains of 1.4 per cent and output gains of 0.8 per cent, prompting comparison with the 'productivity paradox'.

Equity cross-holdings among supply-chain firms are also cited as a structural risk. Because Nvidia and other chipmakers hold equity in many of the same companies that buy their hardware, a slowdown in AI spending could reduce both product revenue and the value of their investment portfolios at once, deepening rather than diversifying exposure. Such minority stakes can also fall outside the regulatory scrutiny that would apply to a full acquisition. By mid-2026 Nvidia's five-year credit-default-swap spread reached a record 82 basis points on 27 July 2026.

Distinctions and scale

On the scale of data-centre financing, Citigroup projected that AI data-centre infrastructure would require $2.8 trillion by 2030, and McKinsey about $7 trillion globally. From January to August 2024, Microsoft, Meta, Google and Amazon spent $125 billion on AI data centres; according to S&P Global, total data-centre market spending was $61 billion in 2025 and data-centre debt issuance $182 billion in the same year. Major technology firms were estimated to spend $650 billion on AI data centres in 2026.

The comparison with the dot-com bubble is a central axis of the discourse. The dot-com bubble peaked on 10 March 2000, when the Nasdaq Composite reached 5,048.62, more than double a year earlier; by October 2002 the Nasdaq-100 had fallen 78 per cent from its peak, and by end-2002 stocks had lost $5 trillion in market value. Just as the name 'dot-com' alone sent share prices soaring, money now crowds into firms whose names contain the keyword 'AI', and investors driven by fear of missing out bet on future growth potential rather than fundamentals. The differences noted are that today's leading firms are large enterprises with actual earnings and cash, and that much of the investment takes the form of physical plant.

The distance from the dot-com bubble is itself contested. Ray Dalio warned that the AI bubble echoes the dot-com in the overvaluation of tech stocks amid low interest rates, and in October 2025 Julien Garran of MacroStrategy Partnership estimated this bubble as 17 times larger than the dot-com bubble and four times larger than the 2008 real-estate bubble. Against that, Mary McNamara argued in late March 2026 that the demise of ChatGPT's Sora text-to-video model was neither 'the first domino [n]or the bursting of the AI bubble'.

An analysis from the standpoint of the rate of profit is also offered. On that view AI investment is large-scale fixed capital investment that raises the organic composition of capital, so that success brings pressure on the rate of profit while failure ends in the destruction of value. Because capital-intensive investment does not greatly increase employment, the question connects to the discussion of jobless growth, and through the cyclical character of chip demand it connects to the semiconductor supercycle.

Examples and related concepts

In the Korean market, the fall of the KOSPI by 44 per cent over 40 days from late June to July 2026, wiping out about $2.18 trillion in market capitalisation, was attributed with Samsung Electronics and SK Hynix accounting for more than half the index to the absence of short-term returns on Big Tech AI infrastructure and concerns over slowing high-bandwidth memory (HBM) demand. The episode is presented as a leading example of an AI bubble burst, but its interpretation is contested.

The structure of the divide between the two camps is also catalogued. The bubble camp cites excessive investment and insufficient profitability (startups such as OpenAI remaining loss-making, with OpenAI burning more than $1 billion a month), circular dealing and self-reinforcement, and overcapacity with weak return on investment and adoption failure. The opposing camp cites that Nvidia, Microsoft and Google already earn tens of billions, that AI infrastructure shows over 90 per cent utilisation from real enterprise customers, and that AI is transformative like electricity or the internet and is a US strategic priority under policies such as the CHIPS Act. The debate remains divided, with some predicting a 2026 collapse and others long-term growth.

Related concepts include AI washing, the cryptocurrency bubble, the Digital Revolution, the Fourth Industrial Revolution, the history of artificial intelligence, the AI winter, the Gartner hype cycle, the productivity paradox, the social media stock bubble, speculation, the workplace impact of artificial intelligence, and the 2025-present global memory supply shortage.