← Steve Eisman hub  ·  Research hub  ·  Research library

Steve Eisman — Inside AI's Fragile Ecosystem with Ed Zitron

"The hyperscalers are investing heavily in GPUs largely to capture revenue from two companies they themselves continue to fund… It is, in Ed's view, something of a closed loop, and both OpenAI and Anthropic remain deeply unprofitable, having lost tens of billions of dollars between them."
2026-SEP-09 · The Real Eisman Playbook (Substack podcast, paid) · Host Steve Eisman · Guest Ed Zitron ("Where's Your Ed At"; "Better Offline") · ↗ Listen / read · captured preview · actionable insights
Paywalled — partial capture. This episode is a PAID Substack post, and Stephen's Chrome profile is not signed in to Substack (the /account warm-up redirected to substack.com/sign-in), so only the free preview rendered: the standfirst plus the complete opening section, "The Circular Economy Supporting Big Tech" (1,413 characters of available content). Everything from the next heading — "Looking Closer at NVIDIA's Numbers" — onward sits behind "This post is for paid subscribers." The audio is preview-only too, so there is no timestamped transcript: the At column links back to the post rather than to a ?t= moment, and every claim on this page comes from the captured preview in transcript.txt. No YouTube upload of this episode was found. Nothing beyond the preview has been inferred or filled in — in particular, the NVIDIA argument that gives the post its second section is not captured, so the NVDA row below carries only the demand-side framing from the first section.
One-line take: Eisman hands the microphone to the AI story's most persistent public critic, and the framing he chooses is the important part — this is not a sceptic being cross-examined, it is a sceptic Eisman says he already reads: "Ed writes thoughtful, deeply researched critiques of the AI story, and I subscribe and read him regularly myself. So I invited him on to hear his full case." The case, as Eisman summarises it in the free section, is a circularity argument rather than a valuation one. The numbers are all revenue-quality numbers: the hyperscalers — Google, Microsoft, Amazon and Meta — "have spent over a trillion dollars in capital expenditure buying GPUs from NVIDIA"; OpenAI and Anthropic "are estimated to make up around 70% of hyperscaler AI revenue"; and UBS projects the two companies could account for 48% of all Google Cloud revenue next year. Put together, "the hyperscalers are investing heavily in GPUs largely to capture revenue from two companies they themselves continue to fund" — a closed loop, with both counterparties "deeply unprofitable, having lost tens of billions of dollars between them." The single sharpest disclosure is on Microsoft: strip OpenAI out and "Microsoft's own AI compute and software business is comparatively modest, a single-digit-billion-dollar business" — because of the $34.3 billion of AI revenue Microsoft reported in fiscal 2026, $24.1 billion came from OpenAI alone. That is the same dependency chain Eisman has been building in his own appearances (Aug-14, Aug-28), but sourced to an outside researcher and pushed one link further: Eisman's version measured hyperscaler exposure to the two labs; Zitron's version says the exposure is self-funded, so the revenue is not independent validation of the capex. What is rowed: only names that actually appear in the captured preview. UBS is cited as the source of a forecast, not discussed as a business, so it gets no row.

1. Stocks & names mentioned

Stance reflects how each name is framed in the captured portion of this post, not a price rating — and the framing is the guest's case as Eisman relays it, which Eisman introduces approvingly but does not explicitly adopt. Rows are limited to companies named in the free preview; the paywalled remainder is not represented. Research: QT Qualtrim · SA Seeking Alpha · STK StockAnalysis.

TickerNameResearchViewWhat was saidAt
AMZNAmazonQT · SA · STK · FANeutralNamed only inside the hyperscaler group — "the hyperscalers, Google, Microsoft, Amazon, and Meta, have spent over a trillion dollars in capital expenditure buying GPUs from NVIDIA" — and so carries the group's exposure: "OpenAI and Anthropic are estimated to make up around 70% of hyperscaler AI revenue." No AWS-specific figure is given in the captured section, and no view on the business itself.read ↗
METAMeta PlatformsQT · SA · STK · FANeutralAlso named only as one of the four hyperscalers inside the trillion-dollar GPU-capex figure. Worth noting that the circularity argument fits Meta least well of the four — it is a buyer of GPUs without a large external AI-cloud business selling back to OpenAI or Anthropic — but the captured section does not draw that distinction.read ↗
NVDANVIDIAQT · SA · STK · FANegativeThe terminal recipient of the loop: the four hyperscalers "have spent over a trillion dollars in capital expenditure buying GPUs from NVIDIA," and the guest's premise is that this spending is being justified by revenue the buyers themselves fund. Zitron is on to argue "why he believes the AI trade is overheated." The post's dedicated section — "Looking Closer at NVIDIA's Numbers" — is behind the paywall and is not captured, so no specific claim about NVIDIA's own accounts is recorded here.read ↗
MSFTMicrosoftQT · SA · STK · FANegativeThe most concrete disclosure in the captured section, and it is a revenue-quality attack: "In fiscal year 2026, Microsoft reported $34.3 billion in AI revenue, and $24.1 billion of that came from OpenAI alone." Net out its own investee and "Microsoft's own AI compute and software business is comparatively modest, a single-digit-billion-dollar business" — i.e. roughly 70% of the headline AI line is money routed back from a company Microsoft funds.read ↗
GOOGLAlphabet (Google)QT · SA · STK · FANegativeThe forward-looking version of the same concentration, and the only sell-side number in the section: "UBS projects the two companies could account for 48% of all Google Cloud revenue next year" — i.e. within twelve months roughly half of Google Cloud's revenue would come from OpenAI and Anthropic, one of which Google itself is a major backer of.read ↗
OpenAIOpenAI (private)NegativeThe single counterparty the loop runs through. It supplies $24.1bn of Microsoft's $34.3bn fiscal-2026 AI revenue, is half of the pair estimated at ~70% of hyperscaler AI revenue and of the UBS 48%-of-Google-Cloud projection, and, with Anthropic, "remain[s] deeply unprofitable, having lost tens of billions of dollars between them" — while being funded by the very hyperscalers booking its payments as revenue.read ↗
AnthropicAnthropic (private)NegativeThe other half of the pair, treated jointly rather than ranked here — no separation into a stronger and weaker lab of the kind Eisman made himself on Aug-28. Both are inside the ~70% of hyperscaler AI revenue and the 48% of Google Cloud next year figures, and both "remain deeply unprofitable."read ↗

2. Talking points

Segmented from the captured free preview only. The post's later sections are paywalled, so this is the opening argument, not the whole episode.

Why this guest — Eisman is a reader, not a sparring partner

The trillion-dollar denominator

Two customers, ~70% of the revenue

The closed loop

Microsoft ex-OpenAI — the number that does the work

Where the post goes next — not captured

3. In plain English

MSFT — Microsoft Negative

Microsoft says it earned $34.3 billion from AI in its 2026 financial year. The argument here is about where that money came from. $24.1 billion of it — about seven dollars in every ten — was paid by OpenAI, a company Microsoft is itself a major investor in and funder of.

Why that matters: normally, revenue is evidence that outside customers want what you sell. If most of it comes from a company you are simultaneously writing cheques to, the revenue is partly your own money coming back around, recorded as a sale. Strip OpenAI out and what is left — Microsoft's AI business selling to everybody else — is described as "a single-digit-billion-dollar business," roughly $10 billion rather than $34 billion.

The practical test this sets up is simple: the part of the AI revenue line to watch is the part that does not come from a company the seller also funds, because that is the only part that proves independent demand exists.

GOOGL — Alphabet (Google) Negative

The same concern, pointed at Google's cloud business and pointed forward rather than backward. The bank UBS projects that next year OpenAI and Anthropic together could be responsible for 48% of all Google Cloud revenue — close to half.

Google is a large backer of Anthropic. So a big slice of what would show up as cloud growth is, again, money flowing out to an AI lab as investment and flowing back in as a bill for computing power. It is not fraudulent or unusual on its own; it just means the growth is not independent confirmation that the world at large is buying AI services.

What to watch: customer concentration. Any business where two customers approach half of a division's revenue is exposed to those two customers' funding — and both of these are private companies that lose money.

NVDA — NVIDIA Negative

NVIDIA sells the chips. The four biggest cloud companies — Google, Microsoft, Amazon and Meta — have spent more than a trillion dollars buying them. That spending is the reason NVIDIA's numbers look the way they do.

The guest's case is that those buyers are spending in order to capture revenue from two loss-making companies they themselves are funding — so the demand behind NVIDIA's sales rests on a financing arrangement rather than on a broad base of profitable customers. If the funding slows, the chip orders slow with it.

An honest caveat on this page: the post's dedicated section on NVIDIA's own figures sits behind the paywall and was not captured, so the detailed argument about NVIDIA specifically is not recorded here — only the demand-side framing above.

OpenAI Negative

OpenAI is the single node the whole chain runs through. It supplies most of Microsoft's reported AI revenue, is half of the pair estimated to account for about 70% of all hyperscaler AI revenue, and is half of the pair UBS thinks could be nearly half of Google Cloud next year.

It is also, together with Anthropic, "deeply unprofitable" — the two have lost tens of billions of dollars between them. The combination is what makes the structure fragile rather than merely unusual: the company generating the revenue that justifies a trillion dollars of chip purchases cannot yet pay for itself, and is kept going by the same companies booking its payments as sales.

Anthropic Negative

Anthropic is the other half of the pair, and in the captured section it is not separated out from OpenAI at all — both are inside the "~70% of hyperscaler AI revenue" estimate, both inside the UBS projection for Google Cloud, and both described as deeply unprofitable.

That joint treatment is itself notable for this archive: in his own interview two weeks earlier Eisman ranked the two explicitly, calling Anthropic the stronger and OpenAI "the weak sister." The guest's framing here does not make that distinction, because the circularity argument does not depend on which lab is healthier — it depends on both being funded by their own customers.


Summary derived from the publicly visible free preview of a paid Substack post (captured text in transcript.txt) for personal study. The remainder of the post is paywalled and is not reproduced or summarised. Not investment advice. © The Real Eisman Playbook / Steve Eisman for source material.