The Real Eisman Playbook Sep 21, 2026 51m 33m saved
With George Noble, former Portfolio Manager at Fidelity, now a newsletter writer and podcaster
About 70% of the hyperscalers' artificial-intelligence revenue comes from two customers, OpenAI and Anthropic, and that is a quarter to a third of their entire cloud business.
Most bear cases on the AI build-out argue about chip demand or about multiples. George Noble's argument is narrower and harder to answer: follow the money back up the chain and almost all of it starts with two companies that lose money.
"Where it starts to get squishy is something like 70% of AI revenue of the hyperscalers is just from OpenAI and Anthropic and that's about 25 to 35% of their entire cloud revenue."
George Noble, a former portfolio manager at Fidelity who now writes a newsletter and hosts a podcast, on The Real Eisman Playbook, has followed Japan for four decades and ran international money through the last two manias, which is why half of this interview is about the yen.
The full interview is covered here so you can skip it. 51 minutes of audio, 18 minutes of reading.
Here are the 8 arguments that matter.
Key Takeaways
70% of the hyperscalers' AI revenue comes from OpenAI and Anthropic, which is 25% to 35% of their whole cloud business — so the chain ends at two loss-making customers
Noble said one of the two, OpenAI, is already in trouble
Five customers are 70% of Nvidia's accounts receivable in a quarter when revenue more than doubled
Japan's 10-year yield just crossed 4% for the first time in 30 years, and on the debt-to-GDP relationship Noble uses it should be nearer 7%
The US lent Japan dollars rather than let it sell Treasuries to defend the yen — a swap line Noble called insanity
Buying back Treasuries against a $40T debt is "trying to kill a whale with a BB gun", and the credibility cost of failing is the real risk
Malinvestment in AI is 24 times the dot-com build-out on Julien Garran's calculation, which Noble cited as his frame for the cycle
Tesla's share price is unchanged over five years, which Noble put down to liquidity and narrative rather than results
Extrapolating Austin's robotaxi rides to the whole US gives about $800M of value, not $800B — and most of Tesla's valuation rests on that business
1. Bessent's BB Gun
The interview opened on Treasury Secretary Scott Bessent's attempt to hold interest rates down, and Noble started by saying which kind of market this is.
This is one of the periods when macro decides what stock picking is worth
It's the quote from Vladimir Lenin where sometimes years go by and nothing happens and sometimes weeks go by and years happen.
George Noble
He said he and the host are both stock pickers, but that there are stretches when the macro is quiescent and stretches when it is a headwind you cannot pick your way through — a homebuilder or a mortgage lender with rates rising is his example of hopeless. He credited the factor framework he uses to Michael Cananter: rates up or down, dollar up or down, economy slowing or accelerating.
On Bessent himself he was sympathetic about the position and unsympathetic about the conduct. Asked a few months ago what advice he would give the Treasury Secretary, he said his answer was to resign, because there is no good move available. The size of the problem is the reason.
The Treasury's buying is too small to matter against $40T
I think there's no question there's upward pressure on rates mostly because of AI, not so much the deficit because the real change is the AI.
George Noble
He compared the buyback program to "trying to kill a whale with a BB gun," and said the second problem is what happens if it fails: if yields go to 5% in two or three weeks, the Treasury Secretary's credibility is gone. The host's view was that this has already happened — rates fell briefly and are meaningfully higher than before Bessent spoke.
Noble's historical parallel was Norman Lamont and the British government's defense of sterling in the early 1990s, which George Soros and Stanley Druckenmiller attacked and profited from. He added a correction to a story that circulates about it: Bessent was a 28-year-old junior at the time and had nothing to do with that trade. The lesson he drew is about when intervention works.
Intervention only works when the fundamentals are on your side
All right, but intervention and stuff like that, you know this, Steve, it only really works when the fundamentals are with you, right?
George Noble
2. The Inflation Regime Turned
Before the AI discussion, both men spent time on why the post-crisis playbook stopped working. Noble's framing, which he attributed to Louis-Vincent Gave, is that the same policy does different things depending on whether the constraint binds: in a deflationary bust, stimulus goes nowhere; with shortages and bottlenecks, the same stimulus produces real inflation.
The market's error flipped direction in 2022
The market you know this Steve the market was constantly up until like 2021 thereabouts. Post GFC the market was constantly overestimating the extent to which inflation would pick up.
George Noble
Since 2022, he said, the market has been consistently underestimating inflation, because fiscal policy now drives the outcome and the constraints bind.
He also dismissed the practice of reading interest-rate forecasts out of futures markets, on the basis of his own undergraduate research at Yale on how well forward markets predict rates.
Forward markets are a bad forecast of interest rates
They're no better than the weatherman.
George Noble
The host's own complaint was about monthly inflation prints. He has no quarrel with Truflation itself, which he described approvingly: "They measure like 10 million different prices and they try to come up with their own estimate of what inflation is." His objection is to the commentators on his monthly calls, who argue over whether the consumer price index came in at 3.8% or 3.6% and miss the regime change underneath.
3. Show Me The ROI
Asked for the short version of the AI debate, Noble gave a one-line test.
The bear case is a demand for returns, not a view on the technology
Show me the ROI.
George Noble
He said the funding is what tells him where this is. Nvidia announcing a $100 billion package for OpenAI is, to him, the whole story.
A supplier financing its customer is the signal
The fact that they're having to do that tells me all I need to know.
George Noble
His framing for the scale of the misallocation came from Julien Garran of MacroStrategy Partnership in the UK, a guest he suggested the host book: Garran calculates the malinvestment at 24 times what the dot-com build-out produced.
He was careful to separate the technology from the equity. The models work, high-powered users will pay for them, and he put the addressable market at perhaps $50 billion to $100 billion. His point is that Chinese models are close enough for most users at a fraction of the cost, and that being right about the technology has never protected the securities. His evidence came from an interview he did with Peter Berezin of BCA Research.
The internet forecasts were right and the stocks still went to zero
Internet traffic compounded at 43% a year for the next 25 years. It went up by like 25 million%. Some crazy number, right? Didn't stop Global Crossing and all the rest of them from going bankrupt.
George Noble
Then the chain itself. Nvidia sells chips, which is fine. Hyperscalers buy chips, which is fine. The problem is who pays the hyperscalers.
The ecosystem rests on two loss-making companies
Where it starts to get squishy is something like 70% of AI revenue of the hyperscalers is just from OpenAI and Anthropic and that's about 25 to 35% of their entire cloud revenue.
George Noble
One of the two, he said, is in trouble, and it is OpenAI. The host agreed and added a timing point of his own: Anthropic needs to get its S-1 filed fast, because what he called token maxing is ending, and the company would rather go public on second-quarter numbers at the peak than on third-quarter numbers after it.
4. The Vendor-Financing Test
Noble then turned the interview around, asking the host to apply his financial-crisis work on accounting to what is happening now. He built the analogy deliberately: a homebuilder whose price appreciation has stalled starts lending buyers the purchase money at 0% to 1% when the market rate is 7%, does not accrue the difference, and books the whole sale as income.
The host's answer separated the mechanism from the verdict.
The related-party structure is not healthy, he said, but it is not automatically fatal — if Anthropic and OpenAI somehow became very profitable, the loans would be good and the structure would not matter. What it is, is a warning not to read the reported numbers at face value.
Circular financing is a reason to look harder, not a verdict on its own
I mean the circular financing kind of drives me crazy but I'll turn it around.
George Noble
He gave a specific example he said almost nobody has picked up. Nvidia's revenue in its most recent quarter was up over 100%, and its filing discloses that 70% of accounts receivable sits with five accounts.
A concentration that size at the top of the market is the thing to look at
I mean, think about it this way that the largest company on planet Earth basically 70% of its revenue comes from five customers is very frightening.
George Noble
Noble's test for the same question is what the lenders are doing.
Banks and private credit are not funding this, and they would if it were good
If there were healthy lenders that thought these are good credits, they'd make loans.
George Noble
They are not, and the host's answer to why was one word: nervous.
5. The Yen Trade, Explained
The host said the dollar-yen trade is his blind spot: "My mental block is the dollar yen trade. I don't understand it." He asked to be taught, which produced the longest stretch of teaching in the episode.
The mechanics first. A Japanese life insurer or a retail saver earns half a percent in the bank and under 1% on government bonds, converts yen into dollars, and buys a US stock or a real-estate investment trust for the yield. If the position rises 10% and the dollar rises 10%, the return is 20% in yen, and for years it worked exactly that way.
The currency then did the opposite. The yen was 80 to the dollar at one point years ago and is now 160.
The yen keeps falling because Japanese rates are too low
The reason the yen keeps weakening is because rates are not high enough.
George Noble
For the size of the gap he pointed listeners at Robin Brooks on X, a former Goldman Sachs economist now at Brookings, whose chart plots government debt against borrowing costs. Switzerland and Sweden sit at one end, Europe in the middle, then the US, then Japan. On that line, Japan is mispriced.
Japan's 10-year yield should be nearer 7% than 4%
It just crossed 4% for the first time in 30 years. It really should be around 7%.
George Noble
The host put the debt ratios alongside it: US government debt at around 125% to 130% of GDP against Japan's 240%, with the US 10-year at 4.8%. He added the standard caveat that much of Japan's debt is held domestically, which nets the ratio down considerably. On where US rates go from here, the conversation noted that the market currently prices the US 10-year rate ten years forward at 6% — a number neither man defended as a forecast, and both treated as the market's verdict on the path.
6. Japan Got A Loan Instead
The reason any of this matters to a US investor is the direction of the flows. Money is fungible, and for two decades an unusual amount of it has come from abroad into US bonds and into the Nasdaq — Japanese savers, and European pension funds with outsized allocations to US technology.
That flow depends on the yield gap, and the gap is closing. A 10-year Japanese government bond at 3% against a US Treasury at 4.8% leaves 180 basis points, where it used to be far more. At some point, the argument runs, Japanese institutions bring the money home, and the reason it matters is their size.
Japan is the world's largest creditor, so its decision is not a domestic story
And the reason this is important is because the Japanese are the world's largest creditor country. They're the savers.
George Noble
Japan is caught between two bad options: let the yen keep weakening and import inflation, or let rates rise and face the same debt-service problem the US has. Defending the currency means selling dollars, and the dollar assets it owns are US Treasuries — which is the part of the story the episode spent most time on. Japan intervened to the tune of what Noble said he thought was $95 billion in a single week.
The threat was implicit and the Treasury heard it
Hey, Scott, it's a really nice bond market you got there. It'd be a shame if something happened to it, right?
George Noble
Rather than let Japan sell Treasuries, the US lent it the money through a swap line, so Japan could buy its own bonds with borrowed dollars. Noble called that insanity. The host, who had said at the start he did not understand what the operation was about, said he did now.
Both then added the same point from different directions: Japan has been selling Treasuries for a while and China has been selling for years, and the US has spent four years giving foreign holders reasons to look elsewhere — assets impounded after the 2022 invasion of Ukraine, tariffs, tax policy. The host's question at the end of it was the obvious one: who buys the bonds.
7. Narrative Dominance
The middle of the conversation was about what has happened to research. Noble said a guest on his own podcast told him that when he started out he assumed people wanted to be told the truth, and learned that they want to be told a story that makes them feel good. He coined a phrase for it on the spot, as a companion to fiscal dominance.
Stories set prices now, not statements
Don't tell me about the fundamentals and the PE and the valuation, all that kind of stuff. Tell me a story.
George Noble
His own experience proved it, he said. He dug his old Fidelity research reports out of boxes in his basement (six or eight pages: strengths and weaknesses, cash flow, income statement, balance sheet, catalyst, insider buying) and wrote his newsletter's first report in that format, cut to three or four pages. On his monthly call he asked his investors what would make the product better, and they told him three or four pages was too long. One page. An elevator pitch. A story.
The host's contribution was the mechanism underneath it: with a large index constituent, price appreciation is amplified by passive buying, so there is no price discovery. He cited Charles Gave's formulation that more money than fools makes a bull market, and more fools than money makes a bear market. Noble's illustration was a consumer lender with a loan-loss reserve he considers far too low trading at six times book, which a traditional financial analyst would find unbearable and the market treats as a growth story.
Both agreed on what would change it.
A real cost of capital is what puts analysts back in business
Well, because when the cost of capital is zero and any schmo can show up and get whatever he wants, pigs fly. Cats sleep with dogs, right? When the cost of capital starts to matter. Oh, so you want to borrow this money? Show us Eisman data centers, right? Show us how you're going to show us the economics of this data center.
George Noble
8. Two Shorts At Scale
The last section was Noble's two short positions, and the thing he wanted noted was their size.
He has never seen shorts this large
I think I've never seen such shorts at scale as these two companies.
George Noble
He has shorted subprime lenders at a $500 million market capitalization; these are trillion-dollar companies.
On SpaceX, he set the revenue multiple aside as already known and said the incremental part of the story is the lockup. The free float went from about 5% of shares to 20% or 25% after the most recent unlock, and roughly 7% more is scheduled to come into the float every few weeks. His point is that the valuation rests on a story plus a very small float, and one of those two things is changing on a published timetable. He added that family offices that bought in at $200 billion, $400 billion and $800 billion valuations, and university endowments whose positions have grown past their diversification limits, are forced sellers on the way through.
He was blunt about the parts of the story that cannot be checked. Predictions set far enough out, asteroid mining being the one he named, cannot be disproved. One that can be, he said, is data centers in orbit.
Space is not uniformly cold, which is the whole problem
Well, but by the way, you talk to any serious physicist, okay, data centers in space, it's ridiculous. Total insanity.
George Noble
So if you're going to build a data center in space, you're gonna have to deal with heat and cold.
George Noble
His own sum of the parts puts Starlink somewhere around $200 billion to $300 billion, the AI business at a fraction of what was paid for it in stock, and the launch business as real but small. In his closing recap the host put the conclusion plainly: "He thinks SpaceX maybe is worth 200 to 300 billion when the market cap is 1.7 trillion."
Tesla is the same trade with a deteriorating business underneath it. He said US sales in July were down 26%, with Europe and China also falling, and that the third quarter faces a brutal comparison because last year's third quarter was inflated by buying ahead of the removal of credits.
Tesla is going cash-flow negative while the car business shrinks
I think US sales in the months of July were down 26%.
George Noble
Capital spending is the other half. Depreciation and capital spending had been running in the same range, he said, on the order of $6 billion to $8 billion; capital spending is going to about $30 billion this year, spent on artificial intelligence, at a company whose return on equity he called abysmal and which he said has never earned its cost of capital.
The valuation, though, is mostly robotaxis, and that is where he did the arithmetic. Austin's service runs in a narrow geofenced area. Take the disclosed number of rides, divide by Austin's population of roughly a million, extrapolate that rate to 350 million Americans, and the robotaxi business is worth about $800 million — not $800 billion. He credited Gordon Johnson with pushing the disclosure that made the calculation possible.
Five years of story has produced no share-price gain
Do you realize Tesla's stock price is unchanged over 5 years? It's done nothing despite all the hype. I think it's all a function of liquidity and narrative.
George Noble
His closing thought was that the index is now carrying both.
Having both in the S&P is an opportunity for a stock picker
And the fact that they managed to jam both of these stocks into the S&P, I'd love to compete as a stock picker.
George Noble
Bonus Insights
His newsletter is 11 for 12, and the one loss is Southwest Airlines
Noble said the only stock he has recommended that has fallen is Southwest Airlines, written up in January: cheap, activist involvement from Elliott Management, the middle seat going away, bag fees arriving, and the observation that earning half of what United or Delta earns would make it a big stock. Oil and the Strait of Hormuz derailed it, and it is down a few percent.
The SpaceX short was published, and it traveled
He said his firm shorted the stock publicly and wrote the first big report on it, which got almost two million views on X in May before the company went public. Asked whether he feels squeezed on the days it rallies, his answer was that it is still meaningfully below where he shorted it, and that he does not mark himself daily.
Another catch line I use, and you can relate to this, others speculate, we invest, right?
George Noble
His sympathy is with the retail investor, not the professionals
He said the shiny object for an individual investor is daily price volatility, which produces the fear of missing out, and that the commentary available to them does not help.
The 1980s Japanese market is his template for what a mania ends like
Japan got up to 60 times earnings. The financials, the banks, maybe this is before your time. Japanese banks were on 100 times earnings, six times, 10 times book.
George Noble
While it was rising, he said, disciplined investors such as John Templeton and Jeremy Grantham were sidelined because they would not own it. Once it turned, avoiding that one block and indexing everything else was enough to win.
Research nobody wants is a leading indicator, in both directions
The host raised David Einhorn saying his model no longer works; Noble matched it with Jeff Vinick, his former Fidelity colleague, who struggled to raise money years ago and found investors were no longer interested in what he did. Both men treated the disappearance of demand for fundamental work as a cycle marker rather than a permanent change, which is why Noble thinks a real cost of capital puts them back in business.
Noble's bottom line is that AI, SpaceX and Tesla are one trade with three tickers: a story funded by circular financing, held up by passive buying and a small float, and priced as though the return on investment has already been demonstrated — which is the thing he says nobody has shown him.
Products, Companies & Tools Mentioned
Nvidia (Announced a $100B package for OpenAI, which Noble treats as the tell; its filing puts 70% of accounts receivable with five accounts in a quarter when revenue more than doubled)
OpenAI and Anthropic (Together about 70% of the hyperscalers' AI revenue and 25% to 35% of their cloud revenue; Noble says OpenAI is the one in trouble, and the host says Anthropic should file its S-1 before token growth slows)
Tesla (Short. US sales down 26% in July, capital spending rising to about $30B, share price unchanged over five years)
SpaceX (Short. The free float has gone from about 5% to 20–25% with roughly 7% more scheduled every few weeks, against a market capitalization the host put at $1.7T)
Starlink (The part of SpaceX Noble says is genuinely worth something — his range is $200B to $300B)
The US Treasury (Buying back debt against a $40T stock, which Noble compared to killing a whale with a BB gun, and lending Japan dollars through a swap line rather than let it sell Treasuries)
Truflation (The host's preferred alternative inflation measure, built from about 10 million prices; his complaint is with the commentary around monthly CPI, not with the data)
Bank of England (The 1992 sterling defense Noble used as the precedent for what happens when officials hold a price the market disagrees with)
Southwest Airlines (The one losing recommendation in his newsletter, with Elliott Management involved, bag fees and the end of open seating as the case)
Global Crossing (His example of being right about internet traffic and still losing everything in the equity)
Books & Resources Mentioned
Robin Brooks (His chart of government debt against borrowing costs is the source for Noble's claim that Japan's 10-year should be nearer 7%)
Julien Garran, MacroStrategy Partnership (Calculates AI malinvestment at 24 times the dot-com build-out; Noble suggested the host book him)
Peter Berezin, BCA Research (Source of the internet-traffic comparison: the bullish 1999 forecasts came true and the stocks still went bankrupt)
Gordon Johnson, GLJ Research (Pushed out the Tesla sales and robotaxi ride data Noble used for his $800M extrapolation)
Stanley Druckenmiller's rebuttal to the Treasury Secretary (Raised as the counter-argument to the official position on rates; Noble said he would take Druckenmiller's side)
Listen to the full episode
🔴 YouTube
Watch the full episode:
If this was worth your time, send it to someone who has to have a view on this.
Get the latest market chatter as it happens: