AI bubble
Last reviewed
May 31, 2026
Sources
20 citations
Review status
Source-backed
Revision
v2 ยท 3,734 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 31, 2026
Sources
20 citations
Review status
Source-backed
Revision
v2 ยท 3,734 words
Add missing citations, update stale details, or suggest a clearer explanation.
The AI bubble is the ongoing debate, most intense across 2025 and into 2026, over whether the surge in artificial intelligence investment, company valuations, and infrastructure spending that began around 2023 amounts to a speculative bubble. The question is contested. Some investors, economists, and even AI executives argue that prices and capital commitments have run well ahead of the revenue and profit that AI products actually generate, which is the classic signature of a bubble. Others argue that AI demand, usage, and revenue are real and growing fast, that the leading companies are highly profitable, and that the infrastructure being built has lasting value even if individual bets fail. As of early 2026 the debate had not been settled, and the same facts were read in opposite directions by serious people on both sides.
This article lays out the bubble thesis, the evidence its proponents cite, the counterarguments, the specific flashpoints of 2025 and 2026, and the historical analogies that get invoked. It attributes claims to their sources and does not take a position on whether a bubble exists or what will happen next.
In finance, a bubble is usually described as a situation in which the price of an asset rises far above any reasonable estimate of its underlying value, driven by speculation, optimism, and fear of missing out, until sentiment reverses and prices fall sharply. The AI bubble thesis applies this idea to the cluster of assets tied to artificial intelligence. That cluster includes the shares of chipmakers and large technology firms, the private valuations of AI model developers, and the enormous sums being spent to build data centers.
Proponents of the thesis generally do not claim that AI is useless or fake. The more careful versions of the argument hold that AI can be a genuinely important technology and still be the subject of a financial bubble at the same time. The claim is about prices and capital flows, not about whether the technology works.
The most concrete piece of evidence is the size of capital expenditure by the large cloud companies, often called hyperscalers. Analysts estimated that combined 2025 capital spending by Microsoft, Alphabet (Google), Amazon, and Meta would land somewhere around 350 to 400 billion dollars, with many estimates clustering near 380 billion after the firms raised their guidance during third-quarter 2025 earnings calls.[1][2] Alphabet lifted its 2025 capex guidance to roughly 91 to 93 billion dollars, Meta guided to about 70 to 72 billion and signaled notably larger spending in 2026, and Amazon guided to well over 100 billion, much of it for its cloud and AI work.[2] Several analysts projected that combined hyperscaler capex could exceed 500 billion dollars in 2026, and Morgan Stanley estimated that cumulative data center spending could approach 3 trillion dollars by 2028.[1][3]
The consulting firm Bain and Company put the challenge in revenue terms. In its 2025 Global Technology Report, Bain estimated that AI companies would need to generate about 2 trillion dollars in combined annual revenue by 2030 to fund the computing power required to meet projected demand, and it projected an annual funding gap of roughly 800 billion dollars between that figure and expected revenue.[4] The estimate assumed something like 200 gigawatts of new data center capacity coming online globally by 2030.[4] Bain framed this as a financing challenge rather than a prediction of collapse.
An earlier and influential version of the revenue-versus-spending argument came from David Cahn, a partner at the venture firm Sequoia Capital, who in 2024 published an essay called "AI's $600 Billion Question." Cahn estimated the gap between the capital being poured into AI infrastructure and the revenue AI applications were actually producing, and asked whether the revenue would materialize to justify the spending.[5] He remained optimistic about AI's long-term potential while urging realism about near-term returns.[5]
A second strand of the bubble argument concerns the structure of the deals that move money among AI companies. Skeptics point to arrangements they describe as circular or round-tripping, in which a small group of interconnected firms invest in and buy from one another. Frequently cited examples include Nvidia's announced plan to invest in OpenAI while OpenAI buys Nvidia chips, OpenAI's large compute deals with Oracle and CoreWeave, and an agreement between OpenAI and AMD that included warrants letting OpenAI acquire AMD shares.[6][7]
The worry is that such deals can make demand look stronger than it is, because the same dollars circulate within the ecosystem and can show up as revenue for more than one party. Critics drew an explicit comparison to the vendor financing of the telecom bubble around 2000, when equipment suppliers such as Lucent and Nortel lent money to customers so those customers could buy the suppliers' gear.[7][8] Coverage in the Financial Times, Wall Street Journal, Bloomberg, and The Economist examined the web of interlocking commitments among Nvidia, OpenAI, Oracle, CoreWeave, AMD, and Microsoft, and the phrase circular financing became common shorthand for it.[6][7]
A third concern is concentration. A historically large share of US stock market value and of recent gains came to rest in a handful of large technology companies, the group often called the Magnificent Seven, which includes Apple, Microsoft, Alphabet, Amazon, Nvidia, Meta, and Tesla.[9] By 2025 these firms represented roughly 35 percent or more of the S&P 500 by market value, among the highest concentrations in decades, and Nvidia alone became the first company to cross a 4 trillion dollar and then a 5 trillion dollar market capitalization.[9] Critics argued that this concentration created a systemic risk, because a reversal in AI sentiment could drag down the broad index funds held by ordinary savers.[9]
The private market drew similar scrutiny. OpenAI reached a valuation around 500 billion dollars in late 2025 through a secondary share sale, after raising a primary round earlier in the year at about 300 billion led by SoftBank.[10] Anthropic raised at valuations that climbed through the year, with reports of rounds around 183 billion and later figures near 350 billion.[10] Elon Musk's xAI also raised at sharply rising valuations.[10] Skeptics noted that these numbers sit far ahead of the companies' current revenue. Private valuations are negotiated in funding rounds and can move quickly, and figures vary by source and date.
A fourth line of evidence concerns whether companies adopting AI are getting their money back. In August 2025 a report from MIT's NANDA initiative, titled "The GenAI Divide: State of AI in Business 2025," drew wide attention for finding that about 95 percent of the organizations it studied were getting zero return on their generative AI investments, while only about 5 percent of custom enterprise pilots reached production and produced measurable value.[11] The report, whose lead author was Aditya Challapally, was based on more than 150 interviews, a survey of around 350 employees, and analysis of roughly 300 public deployments.[11] It distinguished between general tools such as ChatGPT that individuals found useful and bespoke enterprise systems that often failed, and it attributed the failures to a learning gap in how the tools were integrated rather than to the quality of the models.[11]
The finding was cited heavily in bubble coverage and was credited with helping trigger an AI stock pullback in mid to late August 2025.[11] Some analysts cautioned that the 95 percent figure came from a specific sample and reflected the early stage of enterprise adoption rather than a permanent verdict on the technology.
The bear case also gained a prominent public face. In early November 2025, a quarterly regulatory filing for Michael Burry's Scion Asset Management disclosed put option positions against Nvidia and Palantir, covering roughly 1 million Nvidia shares and several million Palantir shares.[12] Put positions in such filings are reported at notional value, which can overstate the actual capital at risk.[12] Burry, who is known for betting against the housing market before the 2008 crash and was portrayed in the film The Big Short, then began posting publicly to criticize what he characterized as aggressive accounting at AI infrastructure companies.[12] His central argument concerned depreciation. He contended that hyperscalers were extending the assumed useful life of AI chips and servers, which lowers annual depreciation expense and thereby flatters reported earnings, and he estimated the effect could understate depreciation by tens of billions of dollars across the industry over the following years.[12] Companies and some analysts pushed back, arguing that longer useful lives reflected genuine hardware improvements and that the accounting was appropriate and audited.[12]
Those who doubt that AI is a bubble, or who think the bubble label is misleading, make several points.
The first is that the revenue is real and growing. Nvidia's data center business grew to an annualized run rate measured in the tens of billions of dollars per quarter, OpenAI's annualized revenue was reported to have climbed into the range of roughly 12 to 13 billion dollars by mid-2025, and consumer products such as ChatGPT reached hundreds of millions of weekly users.[7][13] Defenders argue this is a different situation from many dot-com companies that had little or no revenue.
The second is profitability and balance-sheet strength. The companies leading the AI buildout are among the most profitable in the world and fund much of their capital spending from operating cash flow rather than from speculative equity or debt.[13] Defenders also note that the price-to-earnings multiples of the leading firms in 2025, while elevated, were generally well below the extreme levels seen at the dot-com peak in 2000, when many technology stocks traded at triple-digit multiples or had no earnings at all.[9][13]
The third is that infrastructure has lasting value. On this view, even if some capacity is overbuilt, data centers, power, and chips are durable assets that retain usefulness, much as railways and fiber-optic cable did after earlier booms.[8] The fourth is the productivity argument, that AI may eventually deliver large efficiency gains across the economy that companies will pay for, which would close part of the revenue gap that Bain and Cahn described.[4][5]
Industry figures who pushed back hardest included Alex Karp, the chief executive of Palantir, who repeatedly and forcefully rejected the bubble label through 2025 and argued that his company's accelerating results demonstrated real demand, even as Palantir traded at one of the highest price-to-sales multiples in the S&P 500.[14] On the sell side, analysts such as Dan Ives of Wedbush remained bullish, framing the buildout as a genuine technology shift comparable to the early internet.[7]
A striking feature of the 2025 debate is that several central figures acknowledged bubble-like conditions while continuing to invest. In August 2025, OpenAI chief executive Sam Altman told reporters that investors as a whole were, in his words, overexcited about AI, and he compared the moment to the dot-com era, saying that bubbles form when smart people get overexcited about a kernel of truth that is real.[15] He also said, in substance, that someone was going to lose a great deal of money and someone was going to make a great deal, while maintaining that AI was the most important thing to happen in a long time.[15]
In October 2025, speaking at Italian Tech Week, Amazon founder Jeff Bezos described the environment as an industrial bubble and argued that such a bubble can be good, because the underlying technology is real and society benefits from the resulting investment even when many individual investors lose money.[16] He drew comparisons to the biotech bubble of the 1990s.[16] These remarks were widely reported as examples of leaders conceding bubble dynamics while defending continued spending.
On September 22, 2025, Nvidia and OpenAI announced a letter of intent for a strategic partnership. Under the announcement, Nvidia intends to invest up to 100 billion dollars in OpenAI progressively as capacity is deployed, supporting at least 10 gigawatts of Nvidia systems for OpenAI's next-generation infrastructure, with the first gigawatt targeted for the second half of 2026 on Nvidia's Vera Rubin platform.[17] The arrangement is described in more detail in the article on the Nvidia OpenAI partnership. Sam Altman framed compute as the basis of the future economy, and Nvidia's Jensen Huang described the project's scale in expansive terms.[17] The deal drew immediate scrutiny precisely because of its circular structure, with Nvidia investing cash that OpenAI would in part use to buy Nvidia chips, and it became the most cited single example in the circular financing discussion.[6][7]
The physical buildout itself became a flashpoint. Some spending moved off balance sheets through debt issuance and special purpose vehicles, drawing scrutiny about how the buildout was being financed.[3] Meta raised a large bond offering and used special purpose vehicle structures for data center projects, and Oracle issued substantial debt to fund its expansion.[3] OpenAI's Stargate effort, a multi-year plan to build out large amounts of AI AI infrastructure, was repeatedly cited in discussions of the sheer scale of committed spending.
AI-linked stocks swung sharply during the autumn of 2025. There was a sell-off in August around the MIT report and Altman's comments, and further declines in October and November amid renewed bubble fears, with Nvidia, Palantir, and Oracle among the volatile names.[18] Palantir fell after strong November earnings because of concern about its valuation, and Oracle was volatile after disclosing both a large AI-related backlog and thin cloud margins alongside heavy debt.[18] Despite the swings, major indices remained near record highs for much of late 2025, and many AI stocks recovered from individual drops, so the overall picture as of early 2026 was one of nervousness and sharp swings rather than a sustained crash.[18]
Official bodies weighed in during October 2025. The Bank of England's Financial Policy Committee warned that the risk of a sharp market correction had increased and stated that equity valuations appeared stretched, particularly for technology companies focused on artificial intelligence.[19] The International Monetary Fund, in its Global Financial Stability Report and in comments from Managing Director Kristalina Georgieva, warned of stretched valuations and the possibility of a sharp correction, drawing comparisons to the dot-com period.[20] Federal Reserve Chair Jerome Powell separately described equity prices as fairly highly valued.[20] These bodies generally framed their remarks as financial-stability risk assessments rather than predictions of imminent collapse.[19][20]
The debate leans heavily on comparison to past episodes. The most common is the dot-com bubble of the late 1990s, when speculative investment in internet companies, many without viable business models, crashed in 2000 to 2002. Commentators also point to the telecom and fiber-optic bubble of roughly the same period, when overinvestment in network capacity, based on overestimated traffic growth, left large amounts of unused dark fiber that sat idle for years before demand caught up, and which featured the vendor financing that critics liken to AI circular financing.[7][8] Older still is the British railway mania of the 1840s, a share bubble that ruined many investors yet left behind a rail network that powered the economy for decades, often cited as the archetype of a bubble that destroys capital while building lasting infrastructure.[8]
The economist Carlota Perez, in her 2002 work Technological Revolutions and Financial Capital, described a recurring pattern in which major technologies pass through an installation period marked by speculation and a bubble, followed by a crash, and then a deployment period of broader productive use.[8] Her framework was invoked in 2025 to argue that even a crash would not negate AI's long-term importance.
Those who resist the analogies stress the differences. The companies leading the AI boom are large, established, and highly profitable, with strong cash flows, unlike many dot-com startups, and their valuation multiples, while high, are generally below the 2000 extremes.[9][13] Real revenue and adoption exist and are growing.[7][13] Where the analogies are said to hold are in the concentration of gains, the narrative-driven and fear-of-missing-out character of the investing, the heavy spending on infrastructure that might be overbuilt, and the circular financing reminiscent of the telecom era.[7][8][9] A recurring point made by economists is that a technology can be genuinely revolutionary and still be the subject of a financial bubble, because the railways and the internet were both transformative and the subjects of bubbles that wiped out investors while leaving useful infrastructure behind.[8]
The table below summarizes the main claims on each side as they were presented during 2025 and 2026. It is a map of the argument, not a scorecard.
| Topic | Bubble case | Skeptic-of-the-bubble case |
|---|---|---|
| Capital spending | Hyperscaler capex near 380 billion dollars in 2025 and rising far outruns current AI revenue | Much spending is funded from operating cash flow by profitable firms making long-term bets |
| Revenue gap | Bain estimated a roughly 800 billion dollar annual funding gap toward a 2 trillion dollar 2030 need | AI revenue is growing rapidly and productivity gains could close part of the gap over time |
| Deal structure | Circular financing among Nvidia, OpenAI, Oracle, CoreWeave, and AMD can inflate apparent demand | Such deals are normal strategic partnerships reflecting real expected demand |
| Market structure | Extreme concentration in a few names creates systemic risk if sentiment reverses | Those names have real earnings and dominant positions, unlike many dot-com firms |
| Valuations | Private valuations of OpenAI, Anthropic, and xAI sit far ahead of revenue | Multiples of leading public firms are generally below dot-com peak levels |
| Enterprise ROI | An MIT report found about 95 percent of studied firms got zero return on generative AI | The figure reflects early adoption and a specific sample, not a permanent verdict |
| Infrastructure | Overbuilding risks idle capacity, as with telecom dark fiber around 2000 | Data centers, power, and chips are durable assets that retain value |
| Historical analogy | Resembles dot-com, telecom, and railway bubbles that crashed | The leaders are larger, profitable, and backed by real adoption |
As of early 2026 the question of whether AI investment constitutes a bubble remained open. The factual building blocks were largely agreed upon. Capital spending was very large and rising, the deals among the main players were genuinely interlocked, valuations were high by historical standards, official institutions had flagged stretched prices, at least one widely cited study reported weak near-term enterprise returns, and at the same time AI revenue and usage were growing quickly and the leading firms were profitable. What divided observers was the interpretation. The bubble case read the spending and the deal structures as a speculative excess likely to correct, while the other side read the same figures as a rational, if aggressive, buildout of a real and durable technology. Several prominent figures, including Sam Altman and Jeff Bezos, occupied a middle position, acknowledging froth while continuing to invest. No outcome had been determined, and this article does not forecast one.