'Big Short' Investor Explains How the AI Bubble Will Burst

Summary

• Steve Eisman sees AI’s immediate danger as concentration, not lack of demand. Nvidia grew revenue roughly 110%, yet 70% of its accounts receivable came from five customers; meanwhile, OpenAI and Anthropic reportedly generate 70% of hyperscaler AI revenue, equal to 25-30% of total cloud revenue. “The entire chain from Nvidia to the hyperscalers…rests on the future health and growth of Anthropic and OpenAI.”
• OpenAI looks like the weaker link because its costs are outrunning already-rapid revenue growth. Eisman cited roughly $6.5 billion of quarterly revenue against about $12 billion of costs, with revenue rising $1 billion while costs rose $3 billion; Anthropic’s revenue, by contrast, had grown more than 100% in three months. If OpenAI failed, he argued, “the whole United States of America would go into a recession overnight.”
• The capital loop can hold only while the model companies keep growing and their strategic backers keep funding them. Eisman had assumed venture capital supplied most of the money, but said Amazon, Google, Microsoft, Nvidia and SoftBank are the major funders—creating “a circularity to the financing.” It breaks if either leading lab stumbles and investors stop “ponying up.”
• LLM providers have no durable moat because customers can switch models and are becoming intensely token-cost conscious. After OpenAI and Anthropic raised prices, “token maxing” faded: Uber reportedly exhausted its AI budget in three or four months, while another company spent $500 million before realizing it. Users can reserve premium models for critical work and move routine tasks to open-weight alternatives such as “Kimi K3”: “You don’t need a Cadillac for everything.”
• Eisman suspects existential-risk rhetoric is partly an attempt to manufacture the regulation that would create an otherwise-missing moat. His theory is that incumbents want rules excluding Chinese open-weight models and leaving a protected duopoly. He discounts calls to slow development, arguing that if Anthropic CEO Dario Amodei truly believed his product was dangerous, Anthropic could postpone its IPO; it has not. “I take people seriously when they put their money at risk.”
• Hyperscalers retain scale moats, but AI capex has transformed their cash-flow profile and exposed them to fragile customers. Eisman cited approximately $700 billion of hyperscaler AI spending this year, Oracle’s heavy dependence on OpenAI commitments, and Google raising $85 billion of equity despite historically generating abundant cash. His suggested picks are the recipients of that spending—Nvidia, Micron, GE Vernova, Arista Networks, Cisco and Eaton: “Picks and shovels.”
• For the broader market, Eisman watches the 10-year Treasury rather than the Fed’s overnight rate, with 5% his tentative “Rubicon.” He previously thought 4.5% would trigger trouble and admits, “I was wrong”; with the yield around 4.98% after crossing 5%, he thinks a sustained breakout could produce a correction. On federal debt he wishes the deficit were smaller but doubts an imminent debt spiral because Treasuries underpin the global financial system; the host added that there is currently no alternative.

Deep dive

1. Concentration holds up the AI capital stack

• Eisman’s starting point is customer concentration: an established software company with thousands of customers is resilient, while one with two can be “dead” if either fails. AI increasingly resembles the second company, despite the scale of the spending around it.

• Nvidia’s roughly 110% revenue growth is extraordinary—“the largest company on planet Earth” growing above 100%—but Note 7 of its filing showed 70% of accounts receivable at July-end coming from five customers. The host called that a warning flag, not the end of the world; Eisman used it to illustrate the concentration risk.

• The dependency intensifies downstream. Eisman said OpenAI and Anthropic account for 70% of AI revenue at Microsoft, Amazon and Google, equivalent to 25-30% of their total cloud revenue: “The whole chain” therefore rests on those two labs maintaining growth and paying their commitments.

• OpenAI appears weaker for now, though Eisman stresses that disclosure is inadequate and the ranking could change. He cited roughly $6.5 billion of quarterly revenue, approximately $12 billion of costs, and a three-month period in which revenue rose $1 billion while costs rose $3 billion: “We like to say [that] is upside down.”

2. Circular financing can persist—until confidence disappears

• Eisman originally believed venture capital supplied most model-company funding; he now says the principal backers are Amazon, Google, Microsoft, Nvidia and SoftBank. The host’s pushback—worth keeping—is that vendors appear to fund customers who then make commitments back to those same vendors. “That sounds like a circle.”

• Eisman concedes similarities to the financing loops dramatized in The Big Short, but rejects an automatic collapse call. As long as OpenAI and Anthropic grow “very, very rapidly” and investors keep supplying capital, the chain can hold; the break comes when one lab seriously disappoints and new money retreats.

• Oracle makes the exposure concrete. Its RPO, effectively backlog, reportedly jumped from roughly $150 billion to $400 billion in three months. Eisman said sell-side work attributed about half to OpenAI. He later said Oracle itself was over $600 billion, while still describing roughly 50% of its future revenue or backlog as tied to OpenAI—a company that “bleed[s] money.”

3. LLMs have usage growth but no customer captivity

• “Token maxing” meant employees were instructed to use AI constantly, needed or not. OpenAI and Anthropic initially undercharged relative to token costs, then raised prices; customers became much more cost-conscious.

• The budget shocks were vivid: Uber reportedly exhausted its AI allocation in three or four months, while another company spent $500 million before recognizing the bill. Cost awareness pushed routine workloads toward cheaper open-weight models, leaving premium systems for the highest-value tasks.

• That destroys the loyalty needed for a moat. Developers using Claude today can switch tomorrow if another model is better; leadership rotates among Gemini, ChatGPT and Claude. Unlike Google search at its height, no provider commands habitual, near-unquestioned usage: “There are no moats.”

• Eisman’s explicitly labeled conspiracy theory is that AI-doom rhetoric seeks government intervention. Regulation could bar Chinese models and manufacture a duopoly—a “legal barrier” where no commercial moat existed. Regarding Dario Amodei’s call for a slowdown, Eisman said that if Amodei truly believed his product was dangerous, he could postpone Anthropic’s IPO; it has not been postponed.

4. Hyperscaler strength now finances model-layer weakness

• Hyperscalers do possess moats: almost nobody has “a hundred billion dollars to spend on building data centers.” Their vulnerability is that these capital-heavy, defensible platforms increasingly depend on LLM customers that lack switching costs and remain structurally cash-hungry.

• Microsoft, Amazon and Google once generated more cash than they could reinvest and returned the surplus through buybacks. Data-center spending has now consumed much of that cash flow—sometimes making it negative—and Eisman described Google raising $85 billion in equity as unimaginable only one or two years earlier: “World changed.”

• Oracle shows the market’s changing tolerance. After rising from roughly $230 to $330, the stock later fell through $200 to about $150; even after decent results produced a 7% opening gain, it closed down. The host noted a downgrade and an S&P debt rating of BBB-minus; Eisman said investors were nervous about Oracle’s debt and had no specific news explanation for the decline.

5. Chip depreciation is secondary to whether the chain reverses

• The host raises Michael Burry’s argument that hyperscalers extended chip depreciation lives from roughly three-to-four years to five-to-six, mechanically lowering reported expense even as new generations risk making existing chips obsolete. Eisman grants that the accounting change creates a large profitability benefit.

• His disagreement is practical rather than mathematical: current demand remains so strong that older chips still command demand and higher pricing. “With all due respect to Michael,” the depreciation case is “too academic” for the present market.

• If AI succeeds, changing depreciation schedules will not determine the outcome; if OpenAI fails and commitments reverse, the correction will overwhelm any accounting effect. Burry may be sensing that “something [is] wrong,” Eisman says, but he is pointing to a factor too small to drive the decisive move.

6. Power creates the clearest picks-and-shovels exposure

• On whether US power availability will cap AI deployment, Eisman’s honest answer is: “Get in line. Nobody knows.” Some specialists describe serious delays and others say supply is manageable; he cannot yet reconcile the evidence into a confident bottleneck call.

• GE Vernova, ticker GEV, illustrates the demand shock. A business once assigned roughly negative $3 billion in sum-of-the-parts value now faces US electricity growth of 3-4% annually—incremental demand Eisman compares with adding “two large cities” each year.

• Gas turbines sit at the center of that buildout, alongside the electrical equipment needed to connect new generation. Eisman recalled orders rising about 85%; the machines are enormous and take years to manufacture. The host said GE Vernova’s backlog extends through 2035, making backlog and orders more useful indicators than narrative alone.

• The host’s Memphis example sharpens the physical constraint: when the grid could not serve the Colossus data center quickly enough, 35 portable gas turbines were installed on site. Alternatives such as nuclear and Bloom Energy exist, but Eisman expects most near-term supply to remain gas-turbine-led.

7. Own the spending recipients and demand resilient stories

• Eisman would not invest in an LLM provider because there is no moat, and he is wary of hyperscalers after the erosion of their free cash flow. His suggested picks are the companies receiving the capex: Nvidia, Micron, GE Vernova, Arista Networks, Cisco and electrification supplier Eaton—“picks and shovels” rather than glamorous software.

• The software risk is selective, not universal. The host described a Meta agent she called “Muse” that could book flights and hotels without users visiting travel sites, weakening customer ownership and sidebar advertising; Eisman agreed that payments may also become “dicey.” Deeply embedded enterprise products such as Salesforce or ServiceNow retain stronger switching moats.

• His core method is “story but rational”: he wants real growth tailwinds, dislikes management selling stock and generally avoids cyclicals because that becomes an economic forecast. Moats help explain holdings such as Moody’s and Visa, but are not mandatory. The host said she would not buy Palantir at its current price because it was too expensive, and Eisman agreed.

• On Scott Bessent’s Treasury buyback plan, Eisman said the aim was to buy long-term Treasuries, bring rates down and ease the government’s interest burden. He said the announced $6 billion was negligible relative to a $40 trillion deficit and had failed so far; he suspects another measure may follow but does not know what it will be.

• In macro, the 10-year yield matters more than Fed funds because it determines long-term borrowing costs. Eisman now guesses 5% is the market’s “Rubicon,” after incorrectly choosing 4.5%; by contrast, he doubts an imminent federal debt spiral while Treasuries remain the global financial system’s backbone. He has never owned precious metals because they are unproductive, and regards Bitcoin as worse: it trades opposite the behavior implied by its inflation-hedge thesis, making it “a way to speculate about speculating.”