How OpenAI Could Cause a US Recession | TCAF 264

Summary

• Eisman agrees that an OpenAI failure could trigger a US recession, while saying it would not match the scale of 2008. This is hypothetical: Josh says, “I’m not making that prediction.” Eisman points to concentrated exposure, including five customers accounting for 70% of Nvidia’s accounts receivable and OpenAI representing roughly half of Oracle’s backlog. A host’s separate price-war scenario imagines Oracle going bankrupt.
• The hosts discuss how cheaper open-weight models could pressure AI labs’ prices. Michael says they cost one-tenth as much and compares choosing a model to choosing a car; Josh says enterprise habits may offer Claude some protection, but he would switch if a cheaper model performed almost as well. Eisman says the end of “tokenmaxxing” is starting to affect the story.
• A financing squeeze could strengthen Amazon, Alphabet and Microsoft while limiting independent infrastructure providers. Josh cites CoreWeave’s roughly 15% debt yield and argues corporate and consumer AI demand will persist. Eisman agrees hyperscalers are in a stronger position, while warning that the market depends heavily on two companies.
• The hosts frame AI as a common driver of stocks and Treasury yields. Josh links weaker AI sentiment to the 10-year yield’s fall from 5.4% to 5.1%; Michael agrees the move is about AI, not renewed confidence in government finances. Josh estimates that adding AI infrastructure beneficiaries brings the S&P exposure to about 65%.
• AI’s eventual success does not guarantee today’s leading labs will remain the winners. Eisman invokes first-generation internet companies whose foundations enabled later winners. His “conspiracy theory” is that Anthropic should list before third-quarter results; when asked whether its IPO would follow SpaceX’s path, he says he is unsure.
• Insurance funding and government ownership raise questions about who bears AI investment risk. Michael says policyholders might have received 3% interest where investors would have demanded 12% if they understood the exposure. Eisman says the answer to policyholder risk matters, opposes government equity because it could make companies seem “too big to fail,” and questions whether executives would face personal consequences.
• Housing remains a separate economic weakness that rate cuts alone may not resolve. Michael cites an estimate that total monthly ownership costs are about 50% above five years ago, including insurance, utilities and taxes. Eisman says prices may fall gradually; lower rates would help, but there is “still a long way to go.”

Deep dive

1. OpenAI’s spending commitments create a recession channel

• Eisman distinguishes an industrial failure from a banking collapse. GM’s bankruptcy destroys jobs and suppliers; a JPMorgan failure threatens access to money itself. “If JPMorgan collapses, the Earth will burn up.” An AI collapse could cause substantial economic damage while leaving banks functioning.

• Josh’s OpenAI scenario is explicitly conditional: “I’m not making that prediction,” but he asks whether its bankruptcy tomorrow could immediately cause a US recession. Eisman agrees. In a separate hypothetical, a host imagines a full-scale price war turning Anthropic’s sales growth negative, ending investor funding and forcing investment cuts. That host’s scenario concludes Oracle would go bankrupt because roughly half its backlog comes from OpenAI.

• The connecting mechanism is customer concentration. Eisman cites five customers accounting for 70% of Nvidia’s accounts receivable, which he describes as a proxy for revenue, despite recent quarterly revenue growth above 100%. A host says OpenAI and Anthropic account for 70% of hyperscalers’ AI revenue, with AI representing 25–30% of total cloud revenue. Oracle is more exposed still: roughly half its backlog comes from OpenAI.

2. Cheaper models threaten the labs’ pricing power

• Josh and Michael describe a lack of durable model moats: users can move from Claude to Gemini to GPT. Michael says open-weight models may be worse, but cost one-tenth as much: “you don’t need a Mercedes-Benz every time you go on a trip.” He suggests calls to regulate AI could help incumbents limit competition from open-weight models.

• Josh says his company’s Claude Enterprise subscription has displaced his previous ChatGPT use because employees get accustomed to the tool already being paid for. He sees that as a possible enterprise moat, but says he would switch if an open-weight model performed almost as well at one-tenth the price. He has heard rumors that Claude prices are being lowered.

• Michael describes “tokenmaxxing” as companies ordering developers to use AI around the clock, which became too expensive; he says Uber exhausted its AI budget in three months. Eisman says the end of tokenmaxxing is starting to affect the story. The hosts connect the move from per-seat SaaS to usage-based pricing with greater cost scrutiny: employees choosing cheaper models could pressure OpenAI and Anthropic’s prices.

• Eisman questions OpenAI’s reported annualized revenue moving from $70 billion a month earlier to $50 billion at September-end, but says he does not think the figures are a valid comparison. For Anthropic, he cites $16.8 billion and 43% three-month growth versus 140% in Q2, qualifying, “I could be a bit wrong.” His measurement objection remains: “I have absolutely no idea what annualized sales mean.”

3. Financing stress could hand capacity growth to hyperscalers

• CoreWeave is the financing warning rather than Eisman’s chosen systemic trigger. Josh cites a roughly 15% yield and a 60% price decline; Eisman says the company is spending $640 million on debt repayment, a serious issue given its cash flow. Josh says abandoning a half-built project is not viable; Eisman replies that it could be done, but the company would go bankrupt. He adds, “I don’t think CoreWeave is important enough to shake up the entire AI industry.”

• Josh relays reports from visits to Wall Street banks that AI budgets and related hiring should be larger in 2027 than in 2025 or 2026. He argues that if CoreWeave and Nebius cannot borrow enough to build, Amazon, Alphabet and Microsoft inherit more opportunity. Eisman agrees that hyperscalers’ position has strengthened.

• The hosts discuss how political opposition could slow data center construction. Josh says this would mean fewer computing resources purchased for projects that do not proceed; Eisman adds that demand for existing computing resources would increase. Josh argues that corporate and consumer AI demand will persist. Eisman allows that more companies could emerge to secure computing resources in three to four years. For now, he says, the market is so concentrated that “we can only hope these two companies do well.”

4. The hosts frame AI sentiment as a driver of equities and Treasuries

• After high CPI and PPI figures, interest rates rose, but the cited 10-year Treasury yield fell from 5.4% to 5.1%. Josh attributes the reversal to weakening sentiment around AI; Michael agrees that the move reflects AI trading, not greater confidence in government finances. Josh connects weaker AI sentiment with less bond selling associated with the trade.

• The hosts’ sharp illustration of the preceding market move is Western Digital: down 54%, yet still up 100% for the year. That offers little comfort to peak buyers. Josh also points to Lennar’s rebound and strength in other rate-sensitive stocks as signs of how quickly market leadership can change.

• Eisman resists making valuation the central explanation. “If you tell me how the AI story will turn out, I’ll tell you how the stock market will turn out.” Josh says technology makes up 40% of the S&P, and that Google and Amazon bring the share to about 50%; adding investment banks and infrastructure beneficiaries such as GE Vernova, Quanta, Eaton and Rockwell brings it to about 65%. Josh concludes that “it’s all just one trade. It’s AI.”

5. Insurance funding complicates the question of accountability

• The hosts discuss reporting about insurance money financing AI construction and ask what that could mean for policyholders. In a related discussion of Mark Walter’s insurance company and sports-team holdings, Michael says policyholders would not have received money at 3% if they had understood the exposure; the cost could instead have been 12%. The transcript does not fully clarify the mechanics of that comparison.

• Asked about systemic danger to policyholders, Eisman answers, “I’m not worried. The important thing is finding the answer.” He separately rejects government equity investment in AI companies because it would create an expectation that the recipients were “too big to fail.”

• Eisman supports holding an AI company responsible if it hacks someone or commits wrongdoing. His concern is whether executives who knew about risks would face personal consequences. Invoking the lack of executive punishment after 2008, he asks, “Who would sue me?” Josh, not Eisman, says he expects regulations to come only after a crisis.

6. A winning technology can leave its pioneers behind

• Eisman recalls Henry Blodget correctly predicting the internet’s enormous impact while first-generation companies failed to capture much of the eventual value. AOL, Netscape, Yahoo and Excite are his examples; Amazon is the exception he identifies. Second-generation businesses prospered on foundations built by the first. “Maybe that will happen again this time.”

• His self-described “conspiracy theory” is that Anthropic needs to list before third-quarter earnings, as the end of tokenmaxxing makes its performance look worse. He thinks OpenAI is in a similar situation. The hosts offer SpaceX as an example of dire IPO predictions giving way to typical technology-stock trading; asked whether Anthropic would follow that pattern, Eisman says he is unsure.

7. Housing weakness makes the AI dependence more consequential

• Josh describes housing as the worst market since the financial crisis, while distinguishing the circumstances. Existing owners have been stuck because of high rates; Eisman says a wave of new homes benefited builders including Lennar, Pulte and KB. He says that situation is over as listings return and housing-related stocks fall.

• Affordability involves more than mortgage rates. Michael says insurance, utilities and taxes have risen while house prices have resisted adjustment. He cites an estimate that total monthly homeownership costs are 50% higher than five years ago. Josh says the cost of a loan relative to income may be at its most difficult level in his lifetime.

• Michael points to a roughly 70 million-person age cohort supporting underlying housing demand, but says he cannot afford to buy a home. He expects prices eventually to fall, though gradually over several years. Eisman says the process will be slow; lower rates would help, but the other ownership costs remain a burden.

• Eisman rejects the romance of predicting catastrophe: his own experience brought “tremendous irritation and anxiety.” An AI collapse could produce a severe recession, he says, but the 40-to-1 leverage institutions used in 2008 is absent today. “It probably won’t be on that scale.”