Steve Eisman worries that too much of the AI industry’s spending depends on OpenAI and Anthropic continuing to grow and raise money. If either company could no longer pay for its computing commitments, he says, cloud providers and chip suppliers would also suffer.
In his New Money interview, Eisman discusses that dependence, competition between model companies, the cost of building data centres and the electricity needed to run them. He does not say a crash is inevitable. When asked whether the industry resembles the dot-com bubble, he says it is too early to know.
He also discloses that he owns NVIDIA shares. His concern about the industry’s finances coexists with his interest in companies selling the equipment it needs.
A few customers account for a large share of spending
Eisman starts with a comparison to an established software business. Thousands of customers would spread its risks. If it had only two, losing one could threaten the whole company.
He sees a similar problem in AI. NVIDIA sells chips to large cloud providers. Those providers build data centres and sell computing services to model companies. Eisman argues that some apparently separate sources of demand ultimately depend on the same customers.
He points to Note 7 in NVIDIA’s quarterly filing. At July 26, 2026, five direct customers accounted for 70% of its accounts receivable: money those customers owed NVIDIA at that date. The filing gives individual shares of 22%, 14%, 13%, 11% and 10%.
That figure concerns unpaid customer balances, rather than the share of NVIDIA’s revenue coming from those buyers. The filing also does not name them. NVIDIA’s direct customers include equipment makers, distributors, cloud providers and AI model companies.
Eisman then estimates that OpenAI and Anthropic account for 70% of the AI revenue of Microsoft, Amazon and Google, equivalent to roughly 25% to 30% of their total cloud revenue. Those shares are his interview estimates; the company disclosures reviewed for this article do not establish them.
Oracle’s future sales depend on customers paying
Oracle reports remaining performance obligations, often shortened to RPO. These are contracted sales it has not yet recorded as revenue. Eisman uses the figure as a measure of its backlog of future business.
He describes a sharp rise in that backlog and says it now exceeds $600 billion. Oracle’s September 10 results put the total at $664 billion. Eisman says analysts estimate that around half comes from OpenAI. The results release does not confirm that customer share.
His worry is about the customer’s ability to pay. A large computing commitment is valuable to Oracle only if the customer can fund the work it has agreed to buy.
Eisman regards OpenAI as the weaker of the two major model companies at the time of the interview. He bases that view on reported quarterly revenue and losses, while acknowledging that he does not have enough financial information to be certain and that the position could change.
He cites Wall Street Journal reporting that OpenAI’s quarterly revenue grew more slowly than Anthropic’s, while OpenAI’s losses widened. A public account of that reporting puts OpenAI’s second-quarter revenue at $6.7 billion and its operating loss at $12.3 billion. The interview’s rough description of $12 billion in costs should not be read as the same figure: an operating loss is the shortfall after operating expenses exceed revenue.
The same reporting puts Anthropic’s revenue at $11.6 billion and describes a small operating profit. It also says Anthropic’s adjusted-profit calculation was unclear and had previously excluded stock-based compensation. These are reported private-company figures, rather than a directly comparable set of audited results.
Eisman fears that an OpenAI failure could have effects far beyond the company. He goes as far as suggesting that the United States could fall into recession immediately. That is his scenario, not an outcome established by the financial reports. He says the industry’s customer base may become more diverse over time, but that this will take longer than the companies’ immediate need to remain healthy.
Some suppliers are also investors
For the wider accounting and valuation picture, see AI can be a real technological revolution and still contain a financial bubble.
When the interviewer asks where the model companies get their money, Eisman says he had initially assumed that venture-capital firms supplied most of it. He now points to investments by Amazon, Google, Microsoft, NVIDIA and SoftBank.
The interviewer describes a circle: companies put money into model developers, which then commit to buying computing services from some of those same companies. Eisman agrees that the financing has this feature.
He does not say the arrangement must fail. If OpenAI and Anthropic keep growing quickly and investors continue supplying money, he says, the companies can keep paying for computing and the suppliers can continue expanding.
His concern is what happens if one model company performs badly enough that investors stop putting money in. It may then be unable to pay for its commitments to cloud providers. The providers could reduce their expansion and buy fewer chips and other equipment.
He says the discussion would be different if the two companies accounted for only a small share of their suppliers’ AI sales. His argument depends on the concentration estimates he gives, which remain unverified here.
Eisman doubts model companies can keep customers loyal
Eisman uses the investment term “moat” for a competitive advantage that can last a long time. He gives enterprise software as an example: a product used throughout a company can be difficult to replace because the business depends on it.
He sees a different situation among large language model providers. Customers can choose between models, and the leading product changes. He believes a developer using Claude would switch if a better alternative appeared. That is his view of customer behaviour, rather than evidence that every business could make such a change easily.
He also thinks customers are becoming more careful about AI costs. He describes a period of “token maxing”, when some employers encouraged constant AI use. Tokens are the units of text that models process and generate, and usage is often priced by the number of tokens.
In Eisman’s account, model companies initially charged too little to cover the work, then increased prices. Customers began limiting expensive models to demanding tasks and using cheaper open-weight models for other work. Open-weight models make their trained parameters available for others to run.
He believes competition from those alternatives makes it difficult for model companies to maintain high prices. He distinguishes them from cloud providers, which need enormous amounts of money to build data centres, and from established software companies whose products are deeply embedded in customers’ operations.
His allegation about safety warnings
Eisman also speculates that model companies are exaggerating catastrophic risks to win regulations that would protect them from competition. In his account, restrictions on Chinese models could create a legal barrier that the companies cannot maintain through their products alone.
He argues that calls to slow development sit uneasily alongside plans to raise money from public investors. He presents this as his own theory about the companies’ motives. The interview provides no direct evidence that their safety warnings are fabricated or that they are seeking regulation for that purpose.
Infrastructure spending is using up cash
Eisman says Microsoft, Amazon and Google once generated so much spare cash that they bought back shares because they lacked enough investments to spend it on. He argues that data-centre construction has changed this, leaving much less cash after investment and sometimes a shortfall.
The distinction is between cash generated by the existing business and cash left after building infrastructure. A company can still collect substantial cash from operations while spending more than that on its expansion.
Eisman points to Alphabet’s roughly $85 billion equity financing as something he would not have expected a few years earlier. Alphabet’s June investor presentation described an expected financing total of about that size, alongside plans to spend $180 billion to $190 billion on capital investment during 2026. The company said it would fund investment through operating cash, borrowing and the equity offering.
Oracle’s results show the difference directly. It reported $23 billion of operating cash flow for its fiscal first quarter, but negative $5 billion of free cash flow after infrastructure investment. It was generating cash from its business while spending more on infrastructure than it generated from operations.
Eisman says investors are nervous about how much money companies are committing before they know what the returns will be. He is wary of the large cloud providers for that reason, even while recognising that they have other businesses and advantages over smaller competitors.
He gives less weight to chip depreciation
The interviewer raises Michael Burry’s concern about how companies account for their chips. Spreading the purchase cost over more years reduces the expense recorded each year and makes current profits look higher. If chips become obsolete sooner, the assumed useful life may be too long.
Eisman accepts the accounting point, but thinks demand for older chips remains strong enough that it is not the most important risk at present. In his view, continued rapid growth would make the accounting dispute less consequential. A failure by a major model customer could instead cause a large correction regardless of the depreciation schedule.
He cannot yet judge the electricity constraint
On power, Eisman is less certain. He says he has heard conflicting assessments of whether the United States can supply enough electricity for the planned data centres, and has not been able to resolve them.
He discusses GE Vernova, which sells gas turbines and electrical equipment. He recalls buying its shares after an analyst suggested the company to him. Its prospects had looked poor before rising demand for electricity and data centres changed the outlook, he says.
Eisman describes large turbines that take years to build and buyers trying to reserve production far into the future. He also discusses nuclear power and Bloom Energy as alternatives. Bloom’s data-centre systems use fuel cells, rather than the small turbines described in the interview.
When the interviewer asks whether the long wait for equipment proves that power will become a major obstacle, Eisman repeats that he does not have enough information to know.
Where he would invest, and what he thinks AI could change
Eisman says he would not invest in a model company because he doubts it has a lasting competitive advantage. He would look instead at businesses receiving the infrastructure spending. His examples include NVIDIA, Micron, GE Vernova, Arista Networks, Cisco and Eaton. These are his investment preferences, not recommendations from this article.
He also expects AI agents to affect some existing software businesses. His example is travel booking. If an agent searches different sites and makes the reservation, the traveller may stop visiting a particular site directly. That could weaken the site’s relationship with its customer and reduce the value of advertising shown there.
He presents this as a possibility that is still early, and thinks software deeply embedded in a business may be more resilient.
The interviewer finally asks whether today’s boom will end like the dot-com bubble. Eisman returns to the uncertainty: if OpenAI or Anthropic fails, he expects a major correction and a later generation of companies to pick up the pieces. He does not know whether that will happen.
Sources
- New Money’s interview with Steve Eisman on AI financing, competition and infrastructure
- NVIDIA’s fiscal 2027 second-quarter filing, including Note 7 on customer balances
- Oracle’s fiscal 2027 first-quarter results and contracted future sales
- Alphabet’s June 2026 investor presentation on equity financing and infrastructure spending
- The Next Web’s account of Wall Street Journal reporting on OpenAI and Anthropic’s quarterly finances
- Bloom Energy’s explanation of its fuel cells for data centres