Steve Eisman — "If tomorrow OpenAI failed, the US economy would go into an immediate recession"
"I think it's a little nerve-racking that the entire US economy is dependent upon two companies that lose billions. And I would say between the two, Open AI is the weak sister."
One-line take: the dependency map finally gets numbers on every link, and the conclusion is stated at macro scale rather than portfolio scale. The chain: US GDP grows ~2% this year and "half of it is from AI CapEx"; Nvidia's revenue more than doubled but "70% of its accounts receivable was from five companies"; "Nvidia sells chips to hyperscalers, 70% of hyperscaler AI revenue is from Anthropic and OpenAI, and that equates to about 25 to 35% of their cloud revenue"; "if you look at just Oracle, 50% of its $600 billion backlog is just from OpenAI." Hence: "the entire AI ecosystem food chain is dependent upon the future health and success of Anthropic and OpenAI… that also means that the US economy hinges upon" them — and, put to him at a dinner-party level, "that's 100% accurate." The two labs are now explicitly ranked: Anthropic did ~$11.5B in the June quarter, up over 100% sequentially; OpenAI — "the problem child," "the weak sister" — did ~$6.5B, up only 18%, while costs went to $12.5B, up $3B in three months. "Open AI's revenue in 3 months was up a billion dollars sequentially and its costs were up 3 billion sequentially. That's not the right direction." The clock: "token maxing… ended" around late June/July, so "the third and fourth quarters are going to be much more interesting… than the first half," and the metric is second-half revenue growth, not profitability — "eventually profitability matters, but near-term these stocks don't care." Positioning is on the record and it is not a short: "I am still long Nvidia and several other tech stocks. I'm less long cuz I've gotten nervous. But I am not short," with hedges "possible" and undisclosed. Everywhere else he is a deliberate contrarian against the doom trades: the Treasury's $4B buyback is "a joke," spending the $1T TGA to cap rates is "absurd," QE "had absolutely zero impact on the US economy — the only impact it had was causing stock prices to go up," and on reserve-currency collapse, "Ray Dalio has been talking about this nonsense for the last 20 years… there is no alternative to Treasuries." Banks have "never been this safe in anyone's lifetime"; housing is the one place inflation has done real damage.
Transcript note. At
11:40 the auto-caption renders "
OpenTable I think is in trouble." There is no OpenTable in this conversation — the sentence is about
OpenAI, and it continues "I think the revenue growth has obviously slowed dramatically. It better start reaccelerating soon." Other garbles corrected on this page: "Kcashie"=
Kalshi, "Ben Zaporsky"=
Ben Zipursky (probable; Fordham law professor), "Wolfgang Munchau"=
Wolfgang Münchau, "The Real Investment Playbook" (host)=
The Real Eisman Playbook.
1. Stocks & names mentioned
Stance reflects how each name was framed in this interview, not a price rating. The show is Kalshi-sponsored; the sponsor read is advertising, though Eisman is also asked to react to a Kalshi market, which is treated as content. KKR, Apollo, Blackstone and BlackRock are named once, only as the counterparties to the Nvidia facility he says does not yet exist, and are not given rows. Research: QT Qualtrim · SA Seeking Alpha · STK StockAnalysis.
| Ticker | Name | Research | View | What he said | At |
|---|
| KBWP | Invesco KBW Property & Casualty Insurance ETF | QT · SA · STK · FA | Positive | The one concrete instrument he names, offered as a lower-risk place to sit for someone de-risking out of concentrated AI gains: "I could tell you certain ETFs that you could buy that would be lower risk. Like you could buy the KBWP, which is the property and casualty insurance index ETF. Just as one example." Explicitly an example, not a recommendation. | 14:20 |
| NVDA | NVIDIA | QT · SA · STK · FA | Neutral | Positioning stated flatly against Michael Burry's short: "I am still long Nvidia and several other tech stocks. I'm less long cuz I've gotten nervous. But I am not short." Hedging is "possible" but undisclosed. The concern is the customer list, not the product — "revenue was up more than 100%. However, 70% of its accounts receivable was from five companies." On the $500B third-party capital facility: "that's not exactly" real — per Bloomberg it was still being negotiated and "Jensen went out and basically just announced it anyway… right now there is no $500 billion." | 20:13 |
| MSFT | Microsoft | QT · SA · STK · FA | Neutral | Referenced as part of the hyperscaler class, not named individually in this interview. The sizing that applies to the group: "Nvidia sells chips to hyperscalers. 70% of hyperscaler AI revenue is from Anthropic and OpenAI. And that equates to about 25 to 35% of their cloud revenue." | 4:45 |
| AMZN | Amazon | QT · SA · STK · FA | Neutral | Referenced as part of the hyperscaler class, not named individually here. Carries the same 25–35%-of-cloud-revenue exposure to two loss-making private counterparties that defines the group's risk in this conversation. | 4:45 |
| GOOGL | Alphabet (Google) | QT · SA · STK · FA | Neutral | Referenced as part of the hyperscaler class, not named individually here. Read alongside the Aug 15 disclosure that he sold his position to cut AI exposure — the "still long several other tech stocks" line in this interview does not identify which. | 4:45 |
| Anthropic | Anthropic (private) | — | Neutral | The stronger of the two, for now: June-quarter revenue "was like 11 and 1/2 billion… up over 100% versus the March quarter." But the run is explained away by a demand condition that has since ended — "sometime around late June, July, token maxing… ended" — so "the third and fourth quarters are going to be much more interesting… than the first half." The test is the S-1: "I'm hoping to get some clarity on Anthropic when they put out their S-1. Whether they'll give it or not, I don't know." | 6:42 |
| Kalshi | Kalshi (private, prediction markets) | — | Neutral | The show's sponsor, and also the data source the host puts to him: the market on which lab lists first prices Anthropic at 93%. Eisman engages with the implication — that the S-1 he is waiting on is most likely Anthropic's — rather than with the platform. | 6:16 |
| TSLA | Tesla | QT · SA · STK · FA | Neutral | Raised by the host as the counter-example — years of losses, stock up anyway — to argue profitability may not matter for the AI labs. Eisman half-concedes and redirects: "eventually profitability matters, but near-term these stocks don't care, which is why I'm talking about revenue growth slowing is the more important metric." | 9:13 |
| Situational Awareness | Situational Awareness (private hedge fund) | — | Neutral | The worked example of the one-trade problem: "directionally he's right. It's just a matter of bad risk management" — Eisman: "Totally." The fund was long hardware and short software, which he argues was one position, not a hedged pair: "your longs and your shorts are not independent of one another. They move exactly in unison." Levered 4:1, a temporary reversal "put him out of business really quickly." | 26:10 |
| META | Meta Platforms | QT · SA · STK · FA | Neutral | Mentioned only as the subject of a podcast episode: "Meta just settled this lawsuit which was a California lawsuit which was about their algorithms," the trigger for the June interview with law professor Ben Zipursky on the legal theories behind the algorithm suits. No investment view offered. | 24:30 |
| ORCL | Oracle | QT · SA · STK · FA | Negative | The single most concentrated public exposure in the chain, restated with the number: "if you look at just Oracle, 50% of its $600 billion backlog is just from OpenAI" — half the contracted future revenue owed by the company he calls the weak sister of the two. | 5:14 |
| OpenAI | OpenAI (private) | — | Negative | "The weak sister" / "the problem child," and the sequential math is the case: June-quarter revenue ~$6.5B, "up only 18% versus the March quarter. Worse, its costs went to 12 and 1/2 billion, up 3 billion in 3 months… revenue in 3 months was up a billion dollars sequentially and its costs were up 3 billion sequentially. That's not the right direction." Growth "has obviously slowed dramatically. It better start reaccelerating soon," and with token maxing over "it's going to be harder for these guys to raise capital." The stakes: "if tomorrow OpenAI failed, the US economy I think would go into an immediate recession, and the market would have a massive correction." | 8:33 |
| Bitcoin | Bitcoin (digital asset) | — | Negative | Dismissed in the reserve-currency debate rather than argued about: for Treasuries to be displaced "you need to be as big and as liquid… it's not going to be Chinese bonds, it's not going to be Bitcoin, it's not going to be European bonds." On the recent crypto rally he declines a view entirely — "I have literally no idea. I just don't like to predict markets." | 17:43 |
2. Talking points
0:00 Cold open — the whole thesis in four sentences
- "If tomorrow Open AI failed, the US economy I think would go into an immediate recession. And the market would have a massive correction… I think the whole tech space sells off."
- "I think it's a little nerve-wracking that the entire US economy is dependent upon two companies that lose billions."
3:48 "I don't think it's hidden" — half of GDP growth is one line item
- Asked what he would tell Congress about a looming hidden crisis, he rejects the premise: "I don't think it's hidden. If you look at US GDP growth this year, it's going to be up around 2%. Half of it is from AI CapEx. That's public information… it's not like I discovered the Rosetta Stone."
- The methodological point matters as much as the number: the 2008 work required buying a proprietary loan database; this thesis is assembled entirely from disclosed figures anyone can pull.
4:45 The food chain, link by link
- Nvidia: "revenue was up more than 100%. However, 70% of its accounts receivable was from five companies. Which of those companies, we don't know who they are, but we could all pretty much guess a bunch of them."
- "Nvidia sells chips to hyperscalers. 70% of hyperscaler AI revenue is from Anthropic and OpenAI. And that equates to about 25 to 35% of their cloud revenue."
- "If you look at just Oracle, 50% of its $600 billion backlog is just from OpenAI." Conclusion: "the entire AI ecosystem food chain is dependent upon the future health and success of Anthropic and OpenAI… that also means that the US economy hinges upon" them.
5:49 The two labs are now ranked — OpenAI is "the weak sister"
- First time he separates them explicitly rather than treating "Anthropic and OpenAI" as one exposure: "between the two, Open AI is the weak sister."
- This is a meaningful refinement of the Aug 14–15 framing, where both were named jointly as the thing to monitor.
6:42 Token maxing ended — why H1 numbers do not count
- "I'm pretty sure that Anthropic's revenue story will be good through June. And that's because sometime around late June, July, token maxing — which is where people were just spending, they had no sensitivity to spending, they just spent money like crazy on tokens — token maxing ended. And now people are a lot more self-conscious."
- Therefore "the third and fourth quarters are going to be much more interesting for Anthropic and Open AI than the first half of the year." He is pre-committing to discount a strong first-half print before seeing it.
- The observation point: "I'm hoping to get some clarity on Anthropic when they put out their S-1. Whether they'll give it or not, I don't know." [Host: the Kalshi market prices Anthropic at 93% to list first.]
7:58 The metric, narrowed to one: second-half revenue growth
- Asked to reduce the bubble question to two or three variables: "I think it would be the revenue growth for the second half of this year."
- And what a slowdown transmits into: "I think the whole tech space sells off." July's semiconductor selloff was "the preview," not the event.
8:33 The sequential arithmetic that separates them
- Anthropic: June-quarter revenue "like 11 and 1/2 billion… up over 100% versus the March quarter."
- OpenAI: "around 6 and 1/2 billion in revenue and it was up only 18% versus the March quarter. Worse, its costs went to 12 and 1/2 billion, up 3 billion in 3 months."
- "Simple math: Open AI's revenue in 3 months was up a billion dollars sequentially and its costs were up 3 billion sequentially. That's not the right direction." Note the comparison is quarter-on-quarter, not year-on-year — the whole point is to see the inflection early.
9:13 Does profitability matter? — the Tesla objection
- Host: Tesla was unprofitable for most of its history and the stock went up anyway.
- "Eventually profitability matters, but near-term these stocks don't care, which is why I'm talking about revenue growth slowing is the more important metric." He concedes the premise and uses it to justify the metric he chose.
9:53 The Treasury cannot cap the long end — "it's a joke"
- On the buyback: "has what? $4 billion that they're going to buy Treasury with? It's not even a basis point. It's silly." On signalling: "if your signal is that you're going to spend $4 billion, who cares?"
- On deploying the $1T Treasury General Account: "He's not going to spend a trillion dollars"… "the idea that Treasury should spend a trillion dollars to try and cap long-term rates is absurd."
- The precedent he cites: "the Fed did this with quantitative easing… It had absolutely zero impact on the US economy. The only impact it had was causing stock prices to go up." On the 10-year level itself he refuses a call: "I never make predictions about interest rates."
11:40 Cheap capital is not enough — funding gets harder from here
- Asked whether low rates and available capital keep the train running: "No." [The caption's "OpenTable" here is OpenAI.] "I think the revenue growth has obviously slowed dramatically. It better start reaccelerating soon… I think it's going to be harder for these guys to raise capital. I could be wrong. That's what I think."
- On the $500B Nvidia third-party facility with KKR, Apollo, Blackstone and BlackRock: "that's not exactly" — per Bloomberg the deal was dragging, "so Jensen went out and basically just announced it anyway… right now there is no $500 billion." A live lesson in checking whether an announced financing is executed or aspirational.
13:23 The tax lock — why the selling may not come
- "People have made a lot of money — and I'm not just talking about institutions, I'm talking about individuals as well. They have tremendous unrealized gains."
- "If you sell your Nvidia, you're going to pay 35% plus in capital gains taxes. That is painful… people have such embedded gains that I don't think most people are going to sell."
- The one concrete alternative he offers, framed as an example rather than advice: "certain ETFs that you could buy that would be lower risk. Like you could buy the KBWP, which is the property and casualty insurance index ETF."
15:29 The deficit — 40 years of being wrong should tell you something
- "All things being equal, it'd be better if we had a smaller deficit… [but] I am not of the view that the US economy is going to crash and burn anytime soon because of the deficit."
- "I remember in the '90s Pete Peterson was moaning and groaning… I've been making that argument for 40 years. And when you make an argument for 40 years and you've been wrong for 40 years probably you should ask yourself, okay, why hasn't what I predicted happened?"
- The answer he gives is structural, not moral: "the US Treasury market is the most liquid bond market in the world… the financial system of planet Earth uses Treasuries. The repo market is like a $3 trillion market… It's all T-bills. Nothing else."
17:43 "There is no alternative" — and the rate risk that is real
- The falsification test he offers the host: "Name somebody that you think could be an alternative… it's not going to be Chinese bonds, it's not going to be Bitcoin, it's not going to be European bonds." On Ray Dalio: "talking about this nonsense for the last 20 years… until then I think it's academic."
- But he separates reserve-currency doom from rate risk, and endorses the latter: "part of the reason why the interest rates have gone higher is that the amount of debt being raised for AI is sort of crowding out everything else. So rates could go higher, which would hurt the housing market, it would hurt the economy — but the calamity that some people like to pontificate about, I don't think is realistic."
19:52 Not Burry — long, less long, not short
- Put to him that Michael Burry is short Nvidia and the chip sector: "I am not doing that, by the way, just so you understand. I am still long Nvidia and several other tech stocks. I'm less long cuz I've gotten nervous. But I am not short."
- Hedging: "I might. It's possible… I'd rather keep that to myself for now."
- Timing discipline, stated as the reason not to press: "the big risk to the market, which is huge, is Anthropic and Open AI, but… it could materialize a year from now. And as you know, in markets timing is everything."
20:57 Iran, the K-shape, and "it's all one trade"
- "I don't think the Iran war has real long-term implications for the US economy. It has long-term implications for the United States. But the economy I think as of now is fine."
- "We do have a K-shaped economy. The bottom of the K is struggling. There's no question about that. But as long as AI CapEx keeps powering higher, the economy and the market probably goes higher. It's all one trade."
- On the wealth-effect argument for propping the market up he agrees it is "fair" but denies the mechanism matters: "OpenAI will succeed or fail because it'll succeed or fail. It has absolutely nothing to do with Treasury."
25:42 Situational Awareness and the 1905 carriage trade
- On the blown-up fund: "directionally he's right. It's just a matter of bad risk management" — "Totally."
- The analogy, which is the most transferable idea in the interview: "imagine it's 1905 and you have a thesis that automobiles are going to conquer the world and put all the carriage companies in bankruptcy… you buy every auto and auto parts company and you short every horse carriage company. What you've done is one trade. Your longs and your shorts are not independent of one another. They move exactly in unison."
- "Imagine there's a terrible car accident that gets tremendous amount of press. For a period of time the trade reverses… if you're levered 4:1, you get put out of business really quickly." A correct thesis and a fatal position are not mutually exclusive.
27:39 Banks are safe; housing is where inflation actually bit
- Banks: "post Dodd-Frank [they] were forced to delever enormously… the leverage today is like half of what it was… they may have exposure to private credit, but given how much capital they have, I just don't think it's a systemic issue. The banking system in the United States, I don't think it's ever been this safe in anyone's lifetime."
- Housing: "the area that's been really, really hurt is housing. And it's a lot of hidden costs — the cost of taxes, the cost of property and casualty insurance, the cost of utilities. Because of that inflation there's a whole swath of Americans who've been completely priced out… leaving aside the price of the house, just to bear the annual costs are beyond their means."
- Scope limited deliberately: "that impacts housing-related stocks, but I don't think it really impacts much beyond that." And on redoing the 2006–07 work: "that data is sort of gone… if there was a correction in the housing market, I think it would just be a normal correction. I don't think it would be calamitous."
23:33 The guest roster — and the product schedule
- Guests cited: Wolfgang Münchau on Kaputt: The End of the German Miracle — "why is it that Europe doesn't grow? It just doesn't grow… basically regulation in Europe kills the tech sector in Europe"; law professor Ben Zipursky in June on the legal theories behind the Meta algorithm suits; and Adrian Wooldridge, co-author with Alan Greenspan of Capitalism in America: A History, chosen for the 250th anniversary of the United States.
- The show's self-description: "very eclectic. We do a lot of sell-side analysts, we do authors, we do economists. You get a real education from our podcast."
- The cadence, stated at 30:43: "an interview that goes out on Monday on YouTube, a market wrap which goes out Friday after the close, and on Wednesday on Substack for premium subscribers, an additional interview or masterclass." Upcoming premium masterclass: "how to analyze banks," in a couple of weeks.
3. In plain English
OpenAI Negative
OpenAI builds the models behind ChatGPT. It rents its computing power from other companies, and it is losing money, so it stays alive by raising fresh capital. Eisman now separates it from Anthropic and calls it the weaker of the two — "the weak sister," "the problem child."
The reason is arithmetic done quarter to quarter rather than year to year, which shows a turn much earlier. In the June quarter OpenAI took in roughly $6.5bn, only 18% more than the previous quarter — while its costs rose to $12.5bn, an increase of $3bn in three months. Revenue up $1bn, costs up $3bn: "that's not the right direction." A company whose spending is pulling away from its income has to keep persuading investors to fund the gap, and Eisman thinks that persuasion is about to get harder.
Why anyone outside Silicon Valley should care: OpenAI's promised future spending is other companies' promised future revenue — half of Oracle's $600bn order book, a large share of Microsoft's, and, through the hyperscalers, a share of Nvidia's. Add that AI capital spending is roughly half of all US economic growth this year and the conclusion follows: "if tomorrow OpenAI failed, the US economy I think would go into an immediate recession."
Anthropic Neutral
Anthropic is the other pure model company. On the numbers it is the healthier one: June-quarter revenue of about $11.5bn, more than double the previous quarter.
Eisman's caution is about why that number is so good. Through the first half of the year, he says, customers were "token maxing" — using AI without watching what it cost, the way people spend on a new toy. That behaviour ended around late June or July. So the strong first half measures a spending mood that no longer exists, and the real test is the second half: "the third and fourth quarters are going to be much more interesting for Anthropic and Open AI than the first half of the year."
He is waiting for one specific document. An S-1 is the disclosure a company files before going public, and it would show these revenue and cost figures properly for the first time — "I'm hoping to get some clarity on Anthropic when they put out their S-1. Whether they'll give it or not, I don't know." Prediction markets put Anthropic at 93% to list before OpenAI, so that filing is probably the first real look anyone gets.
NVDA — NVIDIA Neutral
Nvidia makes the chips that AI runs on. Eisman is still a holder and says so plainly, in direct contrast to Michael Burry, who is short the stock: "I am still long Nvidia and several other tech stocks. I'm less long cuz I've gotten nervous. But I am not short."
His worry is not the product but the customer list. Accounts receivable is money already billed but not yet collected, and "70% of its accounts receivable was from five companies." A supplier whose unpaid invoices sit with a handful of buyers inherits those buyers' financial problems — and those buyers, in turn, get much of their own AI revenue from two loss-making labs.
He also throws cold water on a headline: the $500bn financing facility announced with KKR, Apollo, Blackstone and BlackRock is, per Bloomberg, still unsigned. Negotiations were dragging and "Jensen went out and basically just announced it anyway… right now there is no $500 billion." Worth remembering that an announced deal and a completed one are different things.
ORCL — Oracle Negative
Oracle has won enormous contracts to host AI workloads, and reports them as backlog — work sold but not yet delivered or billed. Investors normally treat a big backlog as a promise of future earnings.
Eisman reframes it as a credit exposure to one customer: "50% of its $600 billion backlog is just from OpenAI." Half of what Oracle expects to earn depends on a private, unprofitable company continuing to be able to pay. That makes Oracle the most concentrated public expression of the risk he has been describing — and, usefully for anyone watching, the place where trouble at OpenAI would become visible in a listed share price first.
KBWP — Invesco KBW Property & Casualty Insurance ETF Positive
An ETF is a fund that trades like a share; this one holds property and casualty insurers — the companies that insure homes, cars and businesses. Their earnings come from premiums and investment income, and have essentially nothing to do with whether AI spending continues.
Eisman raises it in answer to a hard practical problem. People who bought the AI winners early are sitting on huge unrealised gains, and selling triggers "35% plus in capital gains taxes." That tax bill keeps them locked into a concentrated position even when they are nervous — which is why he thinks "most people are going to sell" is the wrong assumption to build a crash forecast on. For anyone who does want to reduce risk, he names this ETF as one example of somewhere less exposed to the single trade dominating everything else. He frames it as an illustration, not a recommendation — "just as one example."
TSLA — Tesla Neutral
Tesla appears only as the host's counter-argument: it lost money for years and the share price rose anyway, so perhaps profits do not matter for the AI labs either.
Eisman half-agrees, and the concession is what makes his chosen metric coherent. "Eventually profitability matters, but near-term these stocks don't care — which is why I'm talking about revenue growth slowing is the more important metric." In other words: do not wait for losses to become intolerable, because the market has shown it will tolerate them for a long time. Watch instead for the moment growth decelerates, because that is what a market willing to ignore losses is actually paying for.
Situational Awareness Neutral
Situational Awareness was a hedge fund that blew up in July. It was long AI hardware and short AI software, a structure that looks hedged — a bet on one thing paired with a bet against another.
Eisman's point is that it was not hedged at all, and he explains it with an analogy from 1905: suppose you are sure cars will destroy the horse-carriage business, so you buy every car company and short every carriage company. That feels like two positions. It is one, because both legs depend on the same idea being right at the same time. "Your longs and your shorts are not independent of one another. They move exactly in unison."
The failure mode follows: a temporary scare — "a terrible car accident that gets tremendous amount of press" — reverses both legs together, and with 4:1 leverage the fund is wiped out before the thesis has a chance to be proved right. Being directionally correct is not the same as surviving. The test to run on any long/short book is whether the two sides would move in opposite directions in a shock, or the same one.
Bitcoin Negative
Bitcoin comes up only inside the argument about whether the dollar can lose its central role. Eisman's answer is that Treasuries are irreplaceable for a mechanical reason rather than a patriotic one: the global financial plumbing runs on them. The repo market — where banks lend to each other overnight — is roughly $3 trillion and "it's all T-bills. Nothing else."
To displace that, a rival asset would have to be as large and as easily traded in size. "It's not going to be Chinese bonds, it's not going to be Bitcoin, it's not going to be European bonds." He is not arguing about what Bitcoin is worth — he declines to forecast the recent rally at all — only that it is nowhere near liquid enough to be the alternative that the reserve-currency doom argument requires.
Summary & timestamps derived from the public YouTube video (transcript in transcript.txt) for personal study. Not investment advice. © The David Lin Report / Steve Eisman for source material.