The AI Money Chain, Checked Against the Filings

This is a review, written in our own words, of one podcast interview. It is not a transcript, and it is not a recommendation to buy or sell anything. On October 2, 2026, the YouTube channel New Money published "Big Short Investor Explains How the AI Bubble Will Burst," a conversation with Steve Eisman, the investor Michael Lewis wrote about in The Big Short. The claim that matters for markets is a chain. He says Nvidia, the big cloud companies, and two private AI labs are one loop of commitments. If OpenAI fails, he said, the United States would go into a recession overnight. We checked that loop against the filings and against the labs' reported quarters.

Watch the episode if you want his voice. What follows is our paraphrase of the chain, then the public record. We do not speak for Eisman, for New Money, or for any company named here. A related review, of Ray Dalio's version of an AI bubble, is here.

What he argued

He started with an ordinary software test. A company with thousands of customers can lose one. A company with two customers is in trouble if either one breaks. He said the AI buildout has that shape.

Nvidia, which he said he owns, had just reported revenue growth he put at about 110 percent. He told viewers to open the 10-Q that came out the same night and read Note 7. Seventy percent of accounts receivable at the end of July, he said, came from five customers. He called that a warning flag, and he said it was not the end of the world.

Then the spenders: Google, Meta, Amazon, Microsoft, and Oracle, buying those chips. Leave Oracle aside, he said, and Microsoft, Amazon, and Google get 70 percent of their AI revenue from Anthropic and OpenAI alone. He said that slice equals 25 to 30 percent of their entire cloud revenue.

Oracle, in his telling, shows the same fact as a backlog. Remaining performance obligations jumped when the company reported last year's August quarter, from something like $150 billion to something like $400 billion. About half of that, he said, was OpenAI. The backlog is now over $600 billion, and he still put about half of it on OpenAI, a company that loses money.

The chain, as he drew it, runs from Nvidia through the hyperscalers to Anthropic and OpenAI, because those two create the commitments. If they do not grow, and do not have the cash to pay, the chain slows. He called OpenAI the weaker of the two. He then went further. If OpenAI fails, the whole United States would go into a recession overnight. More customers will show up later. Over the next year or so, he said, these two have to stay healthy.

Asked where the money comes from, he said he had thought it was mostly venture capital, and that this is wrong. Most of it, he said, is Amazon, Google, Microsoft, Nvidia, and SoftBank taking equity. The host called it a circle, money and commitments going out and commitments coming back. Eisman agreed there is a circularity. He loosely compared it to the mortgage years, when one structured product was placed inside the next. His defense of the circle was growth. While the two labs grow very fast and people keep giving them money, the chain holds. If one of them stumbles and the checks stop, the chain does not.

He also set Michael Burry's depreciation argument to one side. Burry has said the clouds stretched the accounting life of chips from three or four years to five or six, which lifts reported profit, and that the chips go obsolete faster than the new schedule. Eisman said older chips are still in demand, so the accounting point is too academic for the moment. If the labs keep growing, the depreciation life will not decide the outcome. If OpenAI fails and the chain runs in reverse, the correction will be about the customer, and the depreciation schedule will be beside the point.

That is the claim this page checks. The rest of the episode covers moats, Chinese open-weight models, electric power, and why he would rather own the companies receiving the cash than the model labs. Those are separate arguments.

What the filings show

Nvidia's quarter ended July 26, 2026. Revenue was $96.221 billion, against $46.743 billion a year earlier. That is a 106 percent increase. His "like 110" was the right neighborhood. The August 26 earnings release put Data Center revenue at $89.0 billion, up 117 percent, GAAP gross margin at 75.0 percent, and net income at $59.7 billion.

Note 7 is where he sent people, and the note says what he said it says. Five direct customers accounted for 22 percent, 14 percent, 13 percent, 11 percent, and 10 percent of accounts receivable. Those shares add to 70 percent. The receivable balance was $63.059 billion, so about $44 billion sat with five names. At the January 25, 2026 year-end it was three names, at 25, 18, and 13 percent, or 56 percent of a smaller balance. The same note says payment is normally due shortly after delivery, and that for some investment-grade customers on large data-center builds Nvidia has offered terms from 90 days out to a year. The company is booking the sale and waiting longer to be paid.

One direct customer was 16 percent of the quarter's revenue, about $15.4 billion, mostly in the Compute & Networking segment. Nvidia also estimates that one AI research and deployment company contributed a meaningful amount of revenue by buying cloud services from Nvidia's customers. That sentence does not name the company. A later contingency does. In August 2026 Nvidia entered guarantees, capped at $105 billion, to support land, power, and buildings with affiliates of SB Energy, on behalf of a customer that is an affiliate of OpenAI Group PBC. The leases cover about 4.25 gigawatts at a campus in Pike County, Ohio. They phase in as construction finishes, with the first phase expected in fiscal 2029, and Nvidia pays if the tenant defaults. That is Nvidia's balance sheet standing behind an OpenAI affiliate's lease. On the same balance sheet, marketable equity securities rose from $12.9 billion at January 25 to $42.8 billion at July 26. The filing does not assign that increase to one lab. It does show a chip company carrying a much larger equity book in the same quarter it insured an OpenAI build.

Oracle's remaining performance obligations were $664 billion on August 31, 2026. The company expects about 13 percent of that, roughly $86 billion, to become revenue within twelve months, 37 percent in the following two years, and 34 percent in years three to five. A year earlier, on August 31, 2025, the backlog was $455 billion. Coverage of the May 31, 2025 figure, the quarter before that jump, puts it near $138 billion. His "150 to 400, then over 600" is that path, rounded. The filed levels are about $138 billion, then $455 billion, then $664 billion.

The half he assigns to OpenAI is not a line in the August 10-Q. The Wall Street Journal reported on September 10, 2025 that OpenAI had agreed to buy roughly $300 billion of computing from Oracle over about five years, starting in 2027. Three hundred billion is 45 percent of $664 billion and 47 percent of the $638 billion Oracle reported at May 31, 2026. That is the neighborhood of "half." S&P's July 2026 review, as described in coverage of the downgrade, treated a single customer as about half of the May book. Oracle's own June 10, 2026 year-end release said something the percentage leaves out: most of the backlog increase in the prior two quarters came from large AI contracts where the customer prepaid for the GPUs or supplied them, and that prepaid and customer-supplied hardware totaled $75 billion. Part of the book is funded up front. Part of it is a promise.

On July 9, 2026, S&P Global Ratings cut Oracle from BBB to BBB-, one notch above junk. Coverage of that cut cited fiscal 2027 capital spending of $90 billion to $95 billion and a widening cash deficit. Fiscal 2026 spending was $55.7 billion against about $32 billion of operating cash flow. The stock jumped the day after the September 2025 backlog print. By September 30, 2026, published quotes put the shares near $138, about 55 percent under the October 2025 peak. He said "150" on October 2. Same rerating, a few dollars apart.

Microsoft printed the cleanest single-customer number in the group. The fiscal 2026 Form 10-K, year ended June 30, records $24.1 billion of revenue from commercial arrangements with OpenAI, including revenue-sharing payments, and $6.0 billion of receivables from OpenAI at year-end. Company revenue was $331.8 billion, so OpenAI was about 7.3 percent of Microsoft. Azure revenue topped $100 billion for the year. The $24.1 billion is about a quarter of that Azure figure, and revenue-sharing inside the $24.1 billion means it is not all rent for GPUs. Commercial remaining performance obligations were $678 billion, up 84 percent. On the July earnings call, the chief financial officer said that backlog grew 25 percent once OpenAI is removed. Earlier, after OpenAI committed to an additional $250 billion of Azure services, OpenAI was about 45 percent of a $625 billion book. Later backlog growth came from other customers, so 45 percent is no longer the standing share. The customer is still the feature of the book. Microsoft also holds an equity-method stake in OpenAI. It is the vendor, the creditor, and an owner.

The 70 percent he used for Microsoft, Amazon, and Google together is a research estimate, and he has said so. On August 12 he wrote that several research firms put the two labs at 70 percent of hyperscaler AI revenue and at 25 to 35 percent of total cloud revenue. "AI revenue" is not a line those three companies report on one page. The public pieces do not share one ratio. At Microsoft, OpenAI alone is about a quarter of Azure. Coverage of UBS work has put the two labs near 27 percent of Google Cloud this year. Coverage of Barclays work has put them near 13 percent of AWS. A blend can land near his 25 to 30 percent of cloud. Each cloud's GAAP revenue is a different exposure. Meta, which he listed among the spenders, is mostly buying chips for its own models. It is a different link from Oracle and Microsoft, which are reselling compute to the labs.

The labs' own numbers are private. The Wall Street Journal reported on August 18, 2026, from people familiar with investor updates, that OpenAI's second-quarter revenue was $6.7 billion, up from $5.7 billion, an 18 percent rise. The operating loss, including stock-based pay, widened from $9.3 billion to $12.3 billion. Sales rose by $1 billion. The loss widened by $3 billion. That is the upside-down quarter he described. He called the $12 billion a cost. The Journal's figure is the loss. The cost base is larger than $12 billion, because $6.7 billion of sales still left a $12.3 billion hole. The scale of the hole is what he had right.

Anthropic, in the same report, more than doubled revenue to $11.6 billion and swung to a small operating profit. His "11 plus" and "over 100 percent" match that account. His "both companies bleed" matches OpenAI. On this quarter, it does not match Anthropic. These figures are what the companies told investors, as relayed by the Journal. They are not audited public filings. They are the best public numbers available, and they are second-hand.

Alphabet shows what happens when the old cash machine no longer covers the build. On June 1, 2026 the company announced equity offerings sized at $80 billion, including a $10 billion private placement to Berkshire Hathaway. Reports the following day put the upsized deal near $85 billion. It is the first equity offering of this kind since the 2004 listing. The stated use includes capital spending on AI infrastructure. The same release put 2026 capital expenditure at $180 billion to $190 billion, and said 2027 spending is expected to rise again. His "$85 billion, in equity" is that deal. Barron's, writing as it was announced, said Alphabet, Microsoft, Amazon, Meta, and Oracle together were set to spend about three-quarters of a trillion dollars on AI data centers in 2026. His "$700 billion" sits in that range. It is an aggregate of guides. It is not one audited line.

Rack wiring: Derrick Coetzee, CC0, via Wikimedia Commons. Hero photograph: Carl Lender, CC BY 2.0, via Wikimedia Commons.

If one lab fails

Take his case. OpenAI is the lab with the widening loss.

Oracle is the sharpest single hit. A reported contract of about $300 billion is on the order of half the backlog the stock was repriced on. The backlog is not cash already collected. Thirteen percent of it is scheduled inside a year. A customer that lost $12.3 billion in a quarter funds a multi-year compute contract by raising money, including from companies that will send it the bill. If that window closes, Oracle is left with buildings sized for a tenant that does not pay. S&P has already moved the rating to BBB- with that risk in the story. The $75 billion of prepaid and customer-supplied hardware, as of the May year-end, means the exposure is smaller than the entire $664 billion. It is large enough to matter for this balance sheet.

Microsoft would lose a customer that was 7.3 percent of last year's revenue and the main reason the contracted backlog grew so fast. Office, Windows, LinkedIn, and the rest of Azure would keep billing. A slower Azure growth rate is a different event from Microsoft failing as a business. The extra $250 billion of Azure services is a promise by OpenAI. A promise from a company that does not fund itself is worth what the next funding round says it is worth.

Nvidia would feel it one step removed, and also on the guarantee. The direct customers are mostly the clouds and the system builders. If Microsoft and Oracle order fewer chips, Nvidia's growth rate slows. A company that just earned $59.7 billion in a quarter, at a 75 percent gross margin, does not go to zero because one indirect buyer stumbles. The $105 billion guarantee is the direct line. It is capped, it phases in over years, and it pays only on a tenant default. It is still insurance written on the customer the chain depends on.

Amazon and Google are tied into this more through Anthropic, and through their own models, than through OpenAI's Oracle contract. One failure leaves the second lab in place. That is why one company and two companies are different tests.

The United States would keep operating the next morning. Current-dollar GDP was $32.563 trillion in the second quarter of 2026, on the September 30 third estimate. OpenAI's quarter, annualized, is under $30 billion, which is under a tenth of one percent of that economy. The large number is the capital spending the labs' promises are used to justify. A spend near $700 billion is a bit over 2 percent of GDP, and only if every dollar is treated as US value added. A large share is imported chips. One lab is not the whole $700 billion. Google, Amazon, and Meta are also building for their own products. Real GDP grew at a 2.2 percent annual rate in that same quarter. A recession is a run of weak quarters. The sentence the arithmetic supports is the quieter one he also used: the chain slows.

If both labs fail

Then his research estimate becomes the stress case. If something like 70 percent of AI-specific cloud revenue, and something like a quarter to a third of cloud revenue at Microsoft, Amazon, and Google, really sits with these two customers, losing both would take the growth layer out of cloud. Cloud is not the whole firm. Alphabet would still have search advertising. Amazon would still have the store. Microsoft would still have the software seat. In his August note he said the hyperscalers have existing franchises, that margins would compress, and that growth would slow. That reading fits the documents better than a liquidation does.

The air would leave first in orders for chips, buildings, and power equipment. Nvidia's five-customer receivable says the chip layer is already owed its money by a handful of buyers, on terms that can run to a year. Those buyers place the next order because the labs, and everyone trying to keep up with the labs, are still expanding. If both labs stop, the next orders shrink, and cash from the sales already booked arrives later than the headline. Power gear, which he discussed later in the episode, would roll over on a longer lag, because those backlogs run toward the end of the decade. A cut of that size would show up in US investment, which added to growth last quarter. A reversal would make growth weaker. The clock on that is orders and funding rounds.

What holds the loop today is new money. OpenAI's loss got $3 billion wider while sales rose $1 billion. Anthropic's reported profit is small next to contracts measured in the hundreds of billions. Amazon, Google, Microsoft, Nvidia, and SoftBank can keep buying stakes, on his account of the funding, and public investors can keep buying the clouds' own new shares, as Alphabet just asked them to. While those checks clear, the labs can pay cloud bills, the clouds can keep building, and Nvidia can keep selling chips to the clouds. He described that circle correctly. The earnings of the two labs do not yet pay for it.

Our comment

The concentration is on paper. Five customers, 70 percent of Nvidia's receivables, plus a $105 billion guarantee that names OpenAI. A reported Oracle contract on the order of $300 billion. A Microsoft revenue line of $24.1 billion, and a backlog whose growth rate drops once that customer is removed. An Alphabet share sale near $85 billion because the cash from the old business no longer covers the build. The podcast put the loop in one sentence. The filings already contained the loop.

The sentence that runs ahead of the evidence is the collapse. One failure, on the public numbers, is an Oracle credit problem, a Microsoft backlog problem, and a slower chip cycle, with Anthropic still standing. Two failures would remove the customers research desks say dominate AI cloud revenue, and they would force a cut in the spending boom. The older businesses of the buyers stay. The stock prices that were paid for the growth are the prices that would have to be rewritten. That is a market event with a real investment slowdown attached to it. It arrives through orders, funding rounds, and credit ratings.

The useful watch list is short. Does the next OpenAI update, or better a filing, show the loss still widening against sales? Does Anthropic's reported profit last more than a quarter? Does Oracle still describe new deals as prepaid, and does the 13 percent of backlog due inside a year actually show up as revenue? Does Microsoft's growth outside OpenAI stay near that 25 percent, or does the ex-OpenAI number roll over? Does Nvidia's receivable share stay at five names and 70 percent, and do the payment terms get longer? Do the equity checks continue at the size of Alphabet's sale? If the checks stop, his chain slows, and the filings already say which balance sheets feel it first.

For the day's tape, see the mid-day note. None of the above is a recommendation to buy or sell any stock, including the ones he said he owns. The episode belongs to its makers. This page is a review of the chain in it.