Could OpenAI Push the US Into Recession? Steve Eisman on Concentration Risk, The Compound and Friends #864

Contents

1. His case, made as strongly as his own evidence allows
2. Standing on the other side: three places the chain doesn’t connect
3. The one condition the whole thing turns on
4. Worth reading alongside
5. 帶得走的一件事

In the capital they favour tall topknots, so a hundred miles out the knots stand a foot high; in the capital they favour broad brows, so out there the paint covers half the forehead.

—— A Chang’an street rhyme, quoted in Book of the Later Han, “Biography of Ma Liao” (Eastern Han; my translation)

On October 16, 2026, in episode 864 of The Compound and Friends, recorded in front of a live audience at the University of Miami, Steve Eisman said something blunt: if OpenAI went bankrupt tomorrow, the US economy would go into recession immediately. His reason is concentration. Nvidia’s latest 10-Q discloses that roughly 70% of its accounts receivable comes from five customers; about 70% of hyperscaler AI revenue comes from Anthropic and OpenAI; and half of Oracle’s backlog is OpenAI alone. The argument has a precondition: a price war in models, and growth at those two labs turning negative. He said himself he doesn’t know whether that happens.

His case, made as strongly as his own evidence allows

Eisman’s first claim is that token maxing is over. He means the wave that ran through this spring: companies ordering staff to use AI around the clock, bills climbing, Uber burning its entire annual AI budget in three months. That was a trial surge, not steady-state demand. Once finance departments looked at the invoices, the sentence changed to “however good the technology is, we obviously can’t spend at this rate.”

He offers two data points. OpenAI said annualized revenue at the end of September was $50 billion; a month earlier the figure quoted was $70 billion. Where did the $20 billion go? Then he added the line that made me laugh: “I have absolutely no idea what annualized revenue means. Nobody does.” On Anthropic, the FT reported that day that quarterly revenue reached $16.8 billion, up 43% in three months, down from 140% growth in the second quarter. From 140 to 43 in one quarter.

Then moats. His verdict is that there are none. He uses Google as the contrast: for years you couldn’t even picture what not using Google would look like. Today you use Claude, dislike something, switch to Gemini, then to ChatGPT, at zero cost. Josh pushed back with the enterprise angle: get every Fortune 500 employee habituated to one model and isn’t that a moat? Eisman’s answer was clean: if an open-weight model does more or less the same thing for a tenth of the price, they’ll use the open-weight model. Then the metaphor I keep thinking about: “You don’t need a Mercedes-Benz for every road trip.”

Only then comes the chain. Nvidia grew revenue over 100% last quarter, which he calls awe-inspiring for the largest company on the planet. But the sentence in the same 10-Q about five customers carrying 70% of receivables is the risk itself. One layer down, those customers are the hyperscalers, whose AI revenue is about 70% Anthropic and OpenAI, with AI running 25% to 30% of their total cloud revenue. One layer below that, half of Oracle’s backlog is OpenAI. Oracle’s chart acted the whole story out: $230 before last year’s third-quarter report, then $330 inside two days when backlog jumped to roughly $400 billion, then back to $230 after Gil Luria’s note pointed out the $300 billion increment was OpenAI alone, and $130 now.

So the chain reads: price war, growth turns negative, funding dries up, the two labs cut their commitments, Oracle’s backlog evaporates, the buildout stops. And he sizes that chain inside the index out loud: tech is 40% of the S&P, add Google and Amazon for 50%, add the investment banks plus GE Vernova, Quanta, Eaton and Rockwell and you’re at 65%. It is one trade, he says, and the trade is AI.

He also separates this from 2008, and that passage is the most useful thing in the episode. What happens if General Motors goes bankrupt? Layoffs, suppliers failing, a regional recession, maybe contagion. When JP Morgan goes down, “planet Earth burns.” The difference is deposits: the government had to step in because people cannot be unable to get their money out of a bank. Data centers are different. “You can’t get your chips out of the data center? That’s bad for you.” This one looks like the General Motors story, he says. Very big, not the end of the world.

Standing on the other side: three places the chain doesn’t connect

Josh made the counter-case on air, and I think he made it better than Eisman answered it.

First, demand shows no sign of softening. The audience in front of them had just spent two days inside every investment bank on Wall Street, and each one has a bigger AI budget for 2027 than for 2026. The consumer side is steadier still: every Google search is already AI, and nobody who uses this is going to announce they’d rather work the way they did in 2022. Hard bond markets change who gets to build capacity, not how many people want the output.

Second, the supply constraint points the other way. Local opposition to data centers is growing, and Eisman grants the buildout will slow a lot. But if new sites don’t get built, existing compute becomes scarcer and more valuable. Josh’s version: if the CoreWeaves and Nebiuses can’t sell debt and can’t supply compute, the ball goes back to the hyperscalers, who now sit in a stronger position than a year ago. Eisman answered “I agree” and steered back to concentration.

Third, the pricing shift arguably makes revenue more honest. For fifteen years enterprise software sold seats — a thousand employees times $80, signed for a year. Now it’s pay for what you use. Eisman reads that as bad news (more expensive, so companies push staff toward cheaper models), but the same fact cuts the other way: usage-based revenue tracks real work, which is closer to the floor of demand than shelf-ware seat licenses ever were. As for 140% falling to 43%, Michael said on the spot that this was never sustainable. That isn’t a rebuttal, but it does warn you against treating an unsustainable base as the start of a trend.

Two more from me. One is definitional: Eisman is right that nobody knows what annualized revenue means, and the knife cuts both ways — using a figure whose definition you reject to prove that $20 billion vanished leaves the inference as loose as the figure. The other is what the market actually did. The night before the show Anthropic’s price on Hyperliquid puked in after-hours and Twitter went jet black; the next morning it was up 5.5%, the ten-year had come from 5.4% to 5.1%, and Lennar — a homebuilder with nothing to do with AI — was the strongest name on the tape. The moment of maximum panic was also the moment the pressure released. And the SpaceX listing, the largest IPO on record, falsified every prediction made about it: it opened, fell, rose, fell back, and traded like an ordinary technology stock.

Eisman left himself exits, too. Asked how dangerous it is for insurers to route policyholder money into data centers: “I don’t know how to answer that.” Asked whether an Anthropic IPO would be well received: “I have no idea.” Asked whether yesterday was the bubble popping: “I honestly don’t know.”

The one condition the whole thing turns on

Both sides hold up. What separates them is one observable condition: how well open-weight models substitute for the work enterprises actually pay for.

Here’s how I read it. The first link in Eisman’s chain is a price war, and a price war requires a good enough substitute. His own phrasing is careful: not as good, but you don’t need a Mercedes for every road trip. That sentence contains the fork. If 80% of a company’s AI workload is cleanup, classification and rewriting, then a model at a tenth of the price takes it, frontier revenue gets squeezed, and his chain connects link by link. If the work companies will pay real money for — code that ships, a full quarter of filings read correctly, no confident fabrication — comes back visibly worse from the cheap model, the price war never starts, the first link never connects, and the concentration discussion has no trigger.

Which runs into a question a lot of people have been asking this week: my holdings have nothing to do with AI, so why are they falling too? The answer is in that 65% he calculated. When one trade is two thirds of the index, index flows stop distinguishing between businesses — your position sits in the bucket labeled risk assets and gets sold with everything else. That doesn’t mean the business got worse. It also doesn’t mean you can ignore it: the number to work out is how much of your holdings’ revenue traces back to those two companies’ capex. That’s computable, no mood reading required.

The second question is why bad news sent prices up. This episode offers a clean sample. Inflation prints ran hot two days in a row, yields should have gone up, and yields went down instead. Michael’s read: the market started believing the flood of AI-related debt issuance would slow, so Treasuries got breathing room. The habit I took from it is this — when two narratives fight for the wheel on the same day, watch which price gives way, and you know which narrative is driving. That day rates gave way, so AI was driving.

The third question is the practical one: should I get out now? Eisman’s answer surprised me. Asked what life was like after being right about subprime, he said it wasn’t fun — a lot of aggravation and tremendous anxiety, people telling him he’d ruined the world and that it was his fault. Then the best line of the hour: the problem with all these people is they all want to be me, and I have news for them, the role of Steve Eisman is taken. He had the glory and doesn’t need it again. There are hundreds of people predicting calamity and one of them will be right and get treated as the person who saw it coming — but, he notes, he put money on it, rather than retweeting bad news.

Worth reading alongside

• The Compound and Friends, episode 864, October 16, 2026, recorded at the University of Miami with Steve Eisman
• The Real Eisman Playbook, Eisman’s own weekly show, launched 2025
• Nvidia’s latest 10-Q, specifically the receivables customer-concentration disclosure
• Gil Luria’s (D.A. Davidson) note on the composition of Oracle’s backlog
• The FT report on Anthropic’s quarterly revenue published the day of the episode
• Michael Lewis, The Big Short, the record of Eisman’s 2008 trade

帶得走的一件事

One idea: to size the risk in anything, start by counting how many people its income comes from.

Every number Eisman cited compresses into that. The software analyst at Oppenheimer who refused to take a fast-growing company public in the 1990s gave one reason: it had two customers, and losing one would end it. She wasn’t looking at the growth rate, she was looking at the denominator. Nvidia’s denominator is five. Hyperscaler AI revenue’s denominator is two. Oracle’s backlog has a denominator of one.

Something I’ve tried, offered rather than assigned: take a sheet of paper, write your top three sources of income, and work out what share the largest one is. Then write what the following week looks like if it disappears tomorrow — concretely, what you do Monday, who gets the first phone call, where the rent comes from. The first time I did it I found my denominator was one, and that I had never thought about who that call would go to.

Raw Markdown

This article is an educational discussion of investment method. It is not advice to buy or sell any individual security, offers no target prices, and does not analyze any current holding. Investing carries risk; make your own decisions or consult a qualified professional.