Eisman says AI bubble rests on two customers

New Money put Steve Eisman, the investor Michael Lewis made famous in The Big Short, on camera to explain how an AI bust could unfold. His answer wasn't about models getting worse.

It was about customers. Or the lack of them.

Eisman, who built his reputation betting against subprime housing before 2008 and now hosts The Real Eisman Playbook podcast with his wife Valerie, said the entire AI buildout carries a concentration risk he called awe inspiring. He walked through a chain that starts at Nvidia, runs through the hyperscalers and ends at two private model labs.

Start at the top. Nvidia just reported revenue growth over 100% and remains the largest company on earth by market cap, which Eisman owns. Then he tells viewers to open the 10-Q to Note 7. As of the end of July, 70% of accounts receivable came from five customers. Warning flag, he said, not the end of the world.

Then the hyperscalers. For Microsoft, Amazon and Google, 70% of AI revenue, which equals 25 to 30% of total cloud revenue, comes solely from Anthropic and OpenAI, he said. Throw in Oracle and the picture sharpens. Oracle's RPO, its backlog, jumped from about $150 billion to $400 billion in three months when it reported the August quarter last October. The stock ripped from $230 to $330 in two days. Sell-side work afterward attributed 50% of that RPO to OpenAI alone. Eisman said Oracle's RPO is now over $600 billion and still about half OpenAI.

That is the setup for what had to be true before his warning could make sense. The hyperscaler AI revenue had to become material to cloud. Oracle had to train investors to cheer RPO as future revenue. Nvidia had to keep recognizing growth while letting receivables concentrate. And the two model labs had to keep signing multi-year commitments that justify $700 billion in hyperscaler AI capex this year, the number Eisman cited for 2026.

What still has to happen is the hard part. Those commitments have to be paid. Eisman pointed to Wall Street Journal reporting that OpenAI did about $6.5 billion in revenue in the second quarter of this year. Anthropic was over $11 billion. But OpenAI had costs around $12 billion in the same quarter, with revenue up $1 billion over three months while costs rose $3 billion. He called it upside down. Anthropic grew over 100% in three months versus about 18% for OpenAI. He labeled OpenAI the weaker entity for now, with the caveat that private numbers are thin.

The money question follows the math. Eisman said he had assumed venture capital funded most of the labs. He now thinks most of the capital comes from Amazon, Google, Microsoft and Nvidia plus SoftBank taking equity stakes, which makes the circle explicit. Money goes out as capacity, then comes back as funding for the customers of that capacity. He said as long as Anthropic and OpenAI keep growing very rapidly and people keep giving them money, the chain holds.

If they do not, he laid out the timeline plainly. Imagine OpenAI fails, he said. It could happen. Then the whole chain slows, and he warned the spillover would be instant and national. He added the industry will eventually diversify beyond two buyers, but that will take time. Within the next year or so, those companies have to stay healthy.

Business risk sits on top of concentration risk. Eisman described a period he called token maxing, when companies told engineers to use AI whether they needed it or not. OpenAI and Anthropic were dramatically undercharging for tokens last year, then raised prices to cover costs. Budgets broke. He cited Uber blowing through its AI budget in three or four months and another firm spending $500 million before it realized it had. Customers got cost conscious and started routing work to open-weight models like Kimi K3 when they do not need a frontier model. His summary was blunt.

There are no moats.

In his telling, Google Search had a moat at 90% share. Salesforce and ServiceNow have moats because switching is brutal. Hyperscalers have a capital moat, no one else can casually spend $100 billion on data centers. But LLM providers do not, because a developer using Claude will switch tomorrow if another model is better. Hyperscalers therefore have moats yet are dependent in cloud on LLM tenants that have none.

That feeds his conspiracy theory about regulation. He argued the recent talk that AI will end the world is manufactured hysteria to invite regulation that would ban Chinese open-weight models and create a legal moat for a U.S. duopoly. He pointed to calls to slow down from Anthropic CEO Dario Amodei and others, then asked why none are postponing IPOs if they truly believe their products are dangerous. He noted Elon Musk's own jab at the dynamic, paraphrased as saying your product will end the world and by the way how much can I allocate to you for the IPO.

Eisman also addressed two lingering pressure points that decide what has to happen next. On power, he agreed with Musk's framing that China has a chip problem and the U.S. has a power problem, and that power is physical infrastructure that takes longer to solve, but said he cannot yet nail down how binding the bottleneck is. On accounting, he engaged with Michael Burry's argument from November that hyperscalers stretched Nvidia chip depreciation from three to four years to five to six years, lifting reported profit. Eisman said demand for older chips remains so strong that prices are up, so for now the change is too academic to drive the outcome. If AI keeps growing, the schedule does not matter. If OpenAI fails and the chain reverses, the correction will have nothing to do with depreciation.

Oracle is already flashing the nerves. After its last print, Eisman noted the stock rose 4 to 5% after hours and 7% at the open then closed down on the day with no news, while S&P has the credit at BBB minus, one notch above junk. Microsoft, Amazon and Google once threw off so much cash they could only buy back stock. Now, he said, cash flow is gone and in some cases negative. Google raising $85 billion in equity, its first since going public, would have sounded insane two years ago.

Per Business Insider, Eisman has separately likened off-balance-sheet financing for AI projects to Enron-era tricks, a comparison not in the New Money sit-down that sharpens his broader skepticism about how the build is funded.

The gap a careful viewer still wants is the missing operating detail from Anthropic and OpenAI that would let anyone test the circle. Commitments, remaining performance obligations by counterparty, and funding terms are mostly inferred from hyperscaler filings and sell-side digging, not from the labs themselves.