“Is the AI bubble about to burst?” – that was the question I asked in July in the pages of “Trud news.” Even then, I noted that the credit market “no longer unconditionally accepts the narrative that every investment in artificial intelligence will inevitably pay off.” The data from the past few months have only confirmed that assessment.
In the third quarter of 2026, OpenAI’s revenue reached approximately $10 billion – nearly 50% more than in the previous quarter. Over the same three months, the company’s adjusted operating loss rose to approximately $10.9 billion. The calculation is simple. For every dollar of reported revenue, there is approximately $1.09 in operating losses. A quarter earlier, the loss had been approximately $1.50 per dollar of revenue.
The ratio is indeed improving. The loss itself, however, continues to grow. As early as the audited 2025 reports, revenue of approximately $13 billion was accompanied by an operating loss of nearly $21 billion. Revenue is growing. Costs are growing along with it, and their rate of growth is often even higher.
Anthropic, OpenAI’s main competitor, is reporting higher revenue. In the third quarter, it approached $17 billion, with an operating loss of $567 million – significantly smaller than OpenAI’s on a comparable basis. At the same time, Anthropic disclosed that it had undertaken commitments worth hundreds of billions of dollars for computing capacity and related infrastructure. According to some regulatory filings and publications, the total value of the company’s existing and expected commitments over the next several years is approximately $900 billion, and may be even higher. The company can hardly cut these costs without consequences. Much of them stem from long-term contracts to reserve computing capacity. If there are not enough customers at the prices factored into the projections to use the reserved capacity, the math no longer works. One quarter with a smaller operating loss (profit is not even a possibility at this stage) does not erase the deep mismatch between today’s customer revenue and the sums the company has already committed to invest.
Both companies have revised their plans to go public – either in terms of timing or conditions. OpenAI’s chief executive made it clear that 2026 was an unsuitable time for a stock-market debut. He referred to the need for decisions on the safety of the systems and the alignment of their behavior with human goals and values. The company’s chief financial officer pointed to 2027 as the more likely timeframe. At the end of October, Anthropic filed a registration statement on Form S-1 with the U.S. Securities and Exchange Commission. In November, the company was expected to begin presenting its shares to potential investors, with an initial valuation of nearly $2 trillion being discussed. Since then, the target valuation has been lowered and, according to reports, is now around $1.4 trillion. Investors are asking questions about the scale of the losses and commitments undertaken, while sell-offs have meanwhile spread across the market as a whole. Potential buyers of the shares are already expressing doubts about whether it is reasonable to follow the original timetable. Such caution reveals how uncertain the companies themselves are that stock-market investors will willingly finance their pace of expansion to date.
The same picture can be seen beyond laboratories specializing in artificial intelligence development. “Microsoft,” “Alphabet,” “Amazon,” “Meta,” and other technology giants operating enormous computing capacity have invested hundreds of billions of dollars in artificial intelligence infrastructure overall.
A number of analyses, including studies cited by the Bank for International Settlements and economists monitoring financial markets, indicate that in certain periods recently, this capital expenditure has exceeded the companies’ operating cash flow. Circular financing is also becoming increasingly visible. One company invests in another or buys from it, while that same company is in turn among its largest customers or suppliers. Once lenders demand more convincing evidence that customer revenue will cover costs, companies can no longer count on receiving the next round of financing by default. That was precisely the change I wrote about in July.
Authoritative observers outside the technology sector are reaching similar conclusions. Ray Dalio, founder of investment firm Bridgewater Associates, sees a classic bubble in the frenzy surrounding artificial intelligence. In his view, the moment is approaching when financing becomes more expensive, holders of paper wealth have to turn it into cash, and the reckoning arrives. “We are at that stage of the cycle which precedes that moment, but we are already getting closer to it,” Dalio says. “I think we are close.”
Michael Burry, who became famous for accurately predicting the 2008 crisis, believes that the artificial intelligence sector is vulnerable to a degree that portends catastrophe. In his words, markets should fall so far that OpenAI and Anthropic will be unable to carry out their initial public offerings on the terms currently under discussion. According to Burry, the two companies will absorb and then destroy trillions of dollars in capital.
Paul Krugman writes that the bursting of the artificial intelligence bubble will almost certainly trigger a recession. Its depth is difficult to predict, but the recovery will probably be slow. Krugman also draws attention to one feature of the current boom. The construction of data centers is displacing other productive investment – a phenomenon that the technological upswing of the late 1990s did not produce.
Dean Baker of the U.S. Center for Economic and Policy Research is tracking the indicators that can show the extent to which market valuations are inflated. He notes that the exact timing of a potential crash is unpredictable, but the consequences would be significant for investment and employment, as well as for household consumption, since household wealth would decline.
Mohamed El-Erian described what is happening as a “rational bubble” that could bring heavy losses to individual investors, even if the technology itself stands the test of time. He draws a direct analogy with the internet bubble of the late 1990s and allows that a crash could follow this time as well. The extent to which the shock spreads through the financial system will also depend on the scale of debt financing. Nothing said so far suggests that the technology itself is useless. The opposite is closer to the truth.
Software that writes code, summarizes documents, creates images, and assists with analysis is already being used for a wide variety of purposes. It will continue to be used. The economic question is narrower and more prosaic. The companies in question raised capital at valuations of hundreds of billions, and in some cases in the range of one to two trillion dollars. Will their customer revenue, after deducting the costs of the infrastructure they have committed to build, be enough to justify these valuations? And will this happen before capital markets refuse them further financing? Judging by the latest quarterly data, the companies are still far from that goal.
Whether the crash comes tomorrow, in six months, or a little later is a secondary question. The direction is what matters. No bubble gives advance warning of the day it will burst. History, however, has repeatedly shown the same sequence – from the railway mania of the 19th century to the internet boom of the late 20th century. First, capital expenditure and company valuations far outpace actual customer revenue. Then new capital becomes more expensive, or there is simply no one left to provide it. After the correction, the technologies that survive are those for which the price customers are willing to pay covers the costs. Some of the capital invested in the belief that, whatever the costs, revenue would eventually catch up with them is lost forever.
The question I raised in July remains valid. The answer is the same.