Steve Eisman on AI in 2026: Why the Big Short Investor Is Long but Hedged

September and October are well covered. Next I’ll pull earlier 2026 episodes and primary transcripts so the chronology isn’t only late-year coverage.
I have the arc from lightening longs to partial hedges. I’ll pull the primary transcripts and fill the January–June gap.
Steve Eisman spent 2026 long the AI buildout while arguing it has one point of failure: OpenAI and Anthropic. He sold Alphabet to cut exposure, refused a directional short for lack of data, then in September put on small “against the box” hedges. The named short he discussed in October was FICO, not an AI stock.

Coverage is dense from late June through October 2, 2026, and thin from January through May. His X account mostly promotes clips; it is not where the thesis lives. No 2026 Bloomberg interview on AI stocks turned up in this search.

How the position actually moved

The mechanism is concentration, not valuation. Eisman’s line is that roughly 70% of hyperscaler AI revenue, and a large share of Nvidia’s receivables, runs through two private labs. If those labs keep growing, depreciation debates and rich multiples do not matter. If one stalls, capex reverses and the suppliers he still likes get hit with everyone else.

That is why he could be “quite long” in August and partially hedged by early October without ever calling a crash. He cut risk by selling a long (Alphabet, around spring 2026, disclosed July 27) and later by shorting part of four unnamed AI holdings he still owns. He has repeatedly said a full short of the boom is premature because the labs are private and there is no delinquency-style data set.

Chronology

January 5, 2026 — The Real Eisman Playbook, Episode 40 (YouTube). With Dan Ives and Chris Verrone. The episode covers “AI 2.0,” Oracle, and whether there are too many AI players. Verbatim Eisman lines on AI stocks were not in the sources retrieved. Treat this as a dated venue, not a quoted call. [1]

January 19, 2026 — The Real Eisman Playbook, Episode 42, Gary Marcus (YouTube). Eisman hosts a critic of LLM scaling. Auto-captions include talk of circular financing and weak return on investment; speaker attribution in those captions is not clean enough to quote as Eisman. [2]

June 1 and June 5, 2026 — his podcast. Episode 62 is titled “Is AI a Bubble? Gary Marcus on What Could Break the Story.” The June 5 Weekly Wrap covers “TokenMaxxing,” Microsoft’s move to token pricing for GitHub Copilot, and OpenAI addiction lawsuits. Titles only; no verified verbatim quotes from this search. [3]

On or before June 22, 2026 — Weekly Wrap; Benzinga write-up June 22. He compared hyperscalers to airlines: capital-hungry, weak pricing power. He kept Nvidia, Arista, and Cisco on the preferred side of the trade and treated Nvidia’s roughly 85% revenue growth as evidence the story was not over. The “story is not over” line is Benzinga’s rendering, not a confirmed verbatim quote. [4]

About June 27–28, 2026 — Fortune interview, as reported by secondary sites. Reported quote: “Even if AI is the greatest thing since the invention of the printing press, there are no moats to shield the providers.” He is described as preferring Nvidia, Arista, and Cisco over Meta, Oracle, Microsoft, and Alphabet, and as calling a SpaceX valuation near 100 times revenue “kind of crazy.” Flag: secondhand. The original Fortune URL was not retrieved. [5] [6]

July 13, 2026 — The Real Eisman Playbook with Torsten Slok. A secondary recap attributes to him: if model companies are giant, capital-hungry, and have no moat, “I’d rather buy Cisco that’s going to supply you,” and notes Oracle’s backlog as roughly half OpenAI. Flag: secondhand paraphrase; not a transcript. [7]

July 27, 2026 — CNBC Squawk Box. This is the first clearly verified de-risking, and it is a sale, not a short.

• “I’ve lightened up.”
• “I sold my Google a couple of months ago… I wanted to reduce my exposure to AI.”
• “People either want to buy AI or they don’t want to buy AI, but they don’t want to shift out of it to buy Clorox.”
• On a hyperscaler cutting capex: “I think the market would go straight down.”
• “It’s questionable whether any of the LLM models have any moats around their businesses. And the Chinese AI models are a lot cheaper.”

He also said the market is “all one trade” and that even a 60/40 portfolio is not diversified because both stocks and new bond issuance are AI-linked. He said he was sitting in cash and had not replaced Google. He did not say he was short AI. [8] [9] [10]

July 27, 2026 — Episode 70, Dan Ives and Gil Luria (same day). “If I was the head of Anthropic or OpenAI, I’d be petrified. That spells to me price war.” He pressed the guests on whether model companies have any moat as Chinese open-weight models undercut U.S. API prices. [11] [12]

August 11–13, 2026 — CNBC Fast Money (full interview page dated August 11; short published August 13). Not shorting. Monitoring.

• “The Achilles heel of this whole story, if there’s going to be an Achilles heel, is if something bad happens to Anthropic and OpenAI.”
• Chinese open models “are much cheaper,” and if they take share “you could have a big price war and then we have a problem.”
• “That’s something I’m monitoring… timing is everything… that story could happen, but it could be a year from now… and it may not happen at all.”
• Until then, “the hyperscalers are going to still be spending money like crazy on Nvidia chips.” [13] [14]

August 14, 2026 — Weekly Wrap, “AI’s Achilles’ Heel” (Benzinga write-up August 17; his own recap clip August 17). Explicitly not short.

• “There is no such data set with respect to AI.”
• Crash calls are “premature.” He remains “quite long.”
• There “don’t seem to be any moats around LLMs.”
• OpenAI and Anthropic are about 70% of AI revenue at Microsoft, Amazon, and Alphabet, and 25% to 35% of cloud revenue; about half of Oracle’s roughly $600 billion backlog is OpenAI. He called that dependency “huge and quite scary.”
• “When OpenAI and Anthropic go public, we will have some real data. Until then, we have supposition.”
• He is not shorting until trouble at the two labs “metastasizes.”
• “Stop rushing to call the top or the bottom.” [15] [16]

August 27, 2026, recorded; published August 28 — The David Lin Report. Hypothetical, not a prediction.

• “If tomorrow OpenAI failed, the U.S. economy, I think, would go into an immediate recession and the market would have a massive correction.”
• “The entire AI ecosystem food chain is dependent upon the future health and success of Anthropic and OpenAI.”
• He said he was not predicting failure, and that the timing was unknown. He has called OpenAI the “weak sister.” [17] [18] [19]

September 4, 2026 — Weekly Wrap, “Is OpenAI the Achilles’ Heel of the US Economy?” He walks the same chain: hyperscaler AI revenue concentrated in the two labs, OpenAI the weaker one, executive departures and slower sequential growth as the tell. Exact English quotes from this episode were not cleanly retrieved. [20]

September 11, 2026 — Weekly Wrap on “Terminator” fears and the 10-year breaking 4.9%. He flags doomsday talk and an OpenAI price cut. The Substack note is paywalled; do not treat secondary summaries as quotes. [21]

September 14, 2026 — Episode 75 with Vincent Daniel and Porter Collins; Benzinga September 15. Eisman, citing Nvidia’s Note 7: “the top five direct customers of Nvidia accounted for 70% of accounts receivable in the quarter.” He also cited reports that “70% of hyperscaler AI revenue is from just Anthropic and OpenAI,” equal to an estimated 25% to 35% of total cloud revenue. Benzinga says he described OpenAI revenue up about $1 billion over three months while costs rose about $3 billion. Porter Collins’s “it scares me” is Collins, not Eisman. [22]

September 17, 2026 — CNBC Squawk Box (about 7:20–7:25 a.m. ET). The viral segment. Still not a short of the AI complex. Benzinga reported he had stopped adding to AI positions and trimmed some exposure; that trimming claim is secondhand relative to the on-air transcript.

• “I think this is all nonsense… All nonsense.”
• “There’s no evidence at all that AI has achieved artificial general intelligence, at all… most of the evidence seems to point that it never will, or if it does, it’s many, many years in the future.”
• “This whole Terminator thing is nonsense.” Benzinga also quotes “garbage. Just garbage.”
• “Token maxing is over. The open-weight models are taking big market share… there are no moats around their business whatsoever, and they’re trying to manufacture a crisis that will create regulation… to create the duopoly that they want.”
• “The entire AI chain, from Nvidia to the hyperscalers to Anthropic and OpenAI, all depends on the future health of Anthropic and OpenAI… if something were to happen to one of those two companies, then the chain would really fall apart. And between the two, OpenAI is the weaker company.”
• “Put your money where your mouth is. Postpone your IPO.” “Is anybody talking about postponing IPOs or postponing raising capital? I haven’t heard anything.” [23] [24] [25] [26]

September 18, 2026 — Weekly Wrap, “Why Dario Amodei and Sam Altman Are Faking the AI Doomsday Crisis.” “There are no pricing moats in this business. Today I have the best LLM and tomorrow yours is better and cheaper.” He argued neither lab can slow down given hyperscaler commitments, and that safety talk is a route to regulation that would wall off open-weight rivals. [27] [28]

September 21, 2026 — Episode 76 with George Noble. Eisman called circular financing something that “kind of drives me crazy,” and Nvidia’s five-customer receivables concentration “very frightening.” When Noble put roughly 70% of hyperscaler AI revenue on OpenAI and Anthropic, Eisman said “Totally agree.” A 24/7 Wall St. piece also attributes “one of them I think is in trouble, which is OpenAI” to this episode. That last sentence is secondhand; the site itself marks awkward wording with “[sic].” [29] [30]

September 25, 2026 — Weekly Wrap on Enron-style AI financing (Benzinga September 28). “Off-balance-sheet techniques are back with a vengeance in the new world of AI. Are we going to do this all over again?” He distinguished open circular financing — Nvidia funding customers out of cash flow, which he said is “all out in the open” — from SPVs and guarantees that keep data-center debt off the balance sheet. “That’s why off-balance-sheet financing is so tempting and has come back in vogue.” [31] [32]

September 28, 2026 — Episode 77 with Cisco IR chief Sam Badri (Benzinga September 29). He said the boom is “for real,” then: “The question is how durable is it?” “You don’t need Mercedes to do every single task.” On a full price war among OpenAI, Anthropic, and Chinese models: “If that were to happen, maybe the story would change. But until then, the story probably continues.” The “for real” wording is Benzinga’s quotation of his reaction to Cisco’s growth, not a full transcript. [33] [34]

October 2, 2026 — Prof G Markets with Ed Elson (YouTube and transcript). This is the partial-hedge interview. The outright short he discusses on the show is FICO, which he says “lives in its own universe” and is not AI-dependent.

• “I’ve played ball too. I’ve invested in these companies. I’ve taken some risk down.”
• “In the past month, I took four of my AI-type stuff, and I just short part of it against the box.” He did not name the four. “I’m not recommending that to anyone.” He told viewers any such hedge has to be unwound by the end of January or the IRS can treat it as a sale.
• “I have some shorts. I’m mostly long. I took down some risk because I think like everybody else, I’m kind of nervous about the whole AI narrative, but I am not willing to make some major call that the whole AI story is going to implode… making that call is premature.”
• “If the AI narrative continues, the stock market will go up. And if the AI narrative breaks, the stock market will have a huge correction… it wouldn’t matter what the valuations are.”
• “If anything bad happens to one of those two companies within the next year… everybody’s in trouble.”
• “I just know I sleep fine.” Contrast with 2008: “I literally thought planet Earth was going to burn.”
• On Anthropic’s reported 2025 numbers (about $4.6 billion of revenue, more than $8 billion operating loss): “I don’t think the 2025 numbers matter.” He floated, and labeled as uncertain, a theory that Anthropic is listing now because the first half of 2026 still looks good from token-maxing, before a second-half slowdown shows up.
• “You’re starting to see signs of a price war breaking out… this business has no moats.”
• If Anthropic failed tomorrow, he said the U.S. economy would be in or near recession quickly, but “that is not systemic risk… it’s not the end of the world.” [35] [35] [36]

October 2, 2026 — New Money, “How the AI Bubble Will Burst” (published that day; record date not confirmed). On Michael Burry’s depreciation thesis: “I think, with all due respect to Michael, I think his argument is too academic.” If the labs grow, a shift from three-to-four-year to five-to-six-year depreciation “is not going to matter.” If OpenAI fails, “we’ll have a massive correction which has nothing to do with the depreciation schedule.” He again put about 70% of Microsoft, Amazon, and Google AI revenue — about 25% to 30% of total cloud revenue — on the two labs, and pointed to Nvidia Note 7: 70% of accounts receivable from five customers as of the end of July. [37] [37]

October 2, 2026 — his Weekly Wrap. Mailbag segment on AI regulation. No verified on-air quote beyond the episode description was retrieved. [38]

Social posts, September 11–22, 2026 — @realsteveeisman. Promotional only: “OpenAI subscriptions,” “Forecasting AI revenues,” “Should AI make us nervous?”, “The AI funding problem.” No standalone thesis. [39] [40]

Not short versus partial hedge

Date	What he said about his book	What it is not
July 27	Sold a long-held Alphabet stake “a couple of months ago” to reduce AI exposure; sitting in cash; no replacement	Not a short. He said defensive rotation (Clorox) does not work.
Aug 11–17	“Quite long.” Crash call “premature.” No data set like Moody’s delinquencies. Monitoring a price war that “could be a year from now” and “may not happen at all.”	Explicitly not shorting the boom.
Aug 27	OpenAI failure would mean “immediate recession” and a “massive correction”	Hypothetical. He said he was not predicting it.
Sep 17	Stopped adding and trimmed, per Benzinga; on air he attacked the doomsday narrative and the lack of moats	Trimming longs is not a short. The “trimmed” detail is secondhand.
Oct 2	“Short part” of four unnamed AI holdings “against the box” over the prior month. “I have some shorts. I’m mostly long.” A major implode call is “premature.”	A hedge of stocks he still owns, not a net short of the AI trade. He said he was not recommending it. The short he walked through in detail was FICO.

“Against the box” means a short paired with an equal long, so the position is neutralized without an outright sale. He warned that if it is not unwound by the end of January, the IRS can treat it as a constructive sale.

Source table

Date	Venue	Position flag	Link	Verification
Jan 5, 2026	Real Eisman Playbook Ep 40	Discusses AI; no verified stock quote	https://www.youtube.com/watch?v=KByibe1S6l8	Venue confirmed; quotes not retrieved
Jan 19, 2026	Ep 42, Gary Marcus	Hosts LLM critic	https://www.youtube.com/watch?v=aI7XknJJC5Q	Auto-captions; speaker unclear
Jun 22, 2026	Weekly Wrap via Benzinga	Bearish on hyperscalers, not Nvidia	https://www.benzinga.com/markets/prediction-markets/26/06/60020062/big-short-star-steve-eisman-turns-bearish-on-hyperscalers-but-not-nvidia	Secondary; some lines paraphrased
~Jun 27, 2026	Fortune, via aggregators	No moats; buy suppliers	https://aiweekly.co/alerts/steve-eisman-picks-nvidia-over-hyperscalers-pans-spacex-ipo	Secondhand. Primary Fortune page not retrieved
Jul 13, 2026	Playbook with Slok	“Buy Cisco” framing	https://www.matterfact.com/newsletter/2026-07-14-ai-capex-tracker-eisman-suppliers	Secondhand
Jul 27, 2026	CNBC Squawk Box	Sold Google; not short	https://www.cnbc.com/video/2026/07/27/big-short-investor-steve-eisman-market-will-go-straight-down-if-any-hyperscaler-cuts-capex.html and https://www.youtube.com/shorts/XoPZ4CJa8WQ	Transcript verified
Jul 27, 2026	Ep 70	“Petrified” / price war	https://www.youtube.com/watch?v=SWEzkOMe4tY	Quote via Benzinga; episode confirmed
Aug 11–13, 2026	CNBC Fast Money	Not short; Achilles’ heel	https://www.cnbc.com/video/2026/08/11/watch-cnbcs-full-interview-with-big-short-investor-steve-eisman.html and https://www.youtube.com/shorts/wiZ6tcOwUkk	Transcript verified
Aug 14, 2026	Weekly Wrap	“Quite long”; not shorting yet	https://www.youtube.com/watch?v=sQCnVoHrN58	Key lines via Benzinga Aug 17: https://www.benzinga.com/markets/prediction-markets/26/08/61246697/steve-eisman-ai-achilles-heel
Aug 27, 2026	David Lin	Hypothetical recession if OpenAI fails	https://www.youtube.com/watch?v=03mCPnemm6M	Quote verified in multiple write-ups; recording date on the video
Sep 14, 2026	Ep 75	70% concentration	https://www.benzinga.com/markets/equities/26/09/61788866/big-short-veterans-say-ai-boom-could-end-in-a-boom-bust-cycle-warn-nvidia-faces-a-hidden-risk-it-scares-me	Benzinga quotes; “scares me” is Collins
Sep 17, 2026	CNBC Squawk Box	Manufacture a crisis; trimmed, per Benzinga	https://www.cnbc.com/video/2026/09/17/big-short-investor-steve-eisman-on-ai-the-companies-are-trying-to-manufacture-a-crisis.html and https://www.youtube.com/watch?v=4qV5WWgFTS8	Core quotes verified; “trimmed” is secondary
Sep 18, 2026	Weekly Wrap	No pricing moats	https://finance.yahoo.com/technology/ai/articles/steve-eisman-says-ai-ceos-033021803.html	Quoted by Benzinga/Yahoo
Sep 21, 2026	Ep 76	“Very frightening”; agrees on 70%	https://finance.yahoo.com/technology/ai/articles/ai-boom-dot-com-subprime-104536006.html	“In trouble, which is OpenAI” is secondhand via 24/7 Wall St.
Sep 25, 2026	Weekly Wrap	Off-balance-sheet warning	https://www.benzinga.com/markets/prediction-markets/26/09/62022808/steve-eisman-ai-off-balance-sheet-financing	Benzinga quotes
Sep 28, 2026	Ep 77, Cisco	Boom “for real,” durability in doubt	https://www.youtube.com/watch?v=Nh4UbFKInwU	“For real” / durability lines via Benzinga
Oct 2, 2026	Prof G Markets	Partial hedge; mostly long	https://www.youtube.com/watch?v=PJrq_bw7bPc and https://podscripts.co/podcasts/prof-g-markets/steve-eisman-one-company-could-break-the-ai-boom	Transcript verified
Oct 2, 2026	New Money	Burry “too academic”	https://www.youtube.com/watch?v=Zw0I7kb6zg4	Auto-transcript; record date not confirmed
Sep–Oct 2026	@realsteveeisman	Promo clips only	https://x.com/realsteveeisman	Not original analysis

What this means if you are trying to trade his words

He never, in the verified record, said he was short the AI boom. The October hedge is a reduction of four longs he would not name, paired with an explicit refusal to call an implosion. The short he was willing to explain was FICO.

The tradable tell he keeps repeating is not the Cape ratio. It is whether open-weight models keep taking share, whether Anthropic’s S-1 shows a second-half 2026 slowdown versus the first half, and whether OpenAI or Anthropic stumbles badly enough that hyperscalers cut capex. He said that last event would send the market “straight down,” and that until it happens the supplier story “probably continues.”

Gaps that would still change this chronology: a primary Fortune page for the late-June interview, full transcripts of the January–June Weekly Wraps, and any Bloomberg hit in 2026. None of those were retrieved.

---

Recent Findings Supplement (October 2026)

Steve Eisman's public commentary on the AI boom from mid-2026 onward has centered on its real underlying demand (evidenced by infrastructure spending and revenue growth at companies like Cisco) alongside acute concentration risks in OpenAI and Anthropic, skepticism toward AI safety/doomsday narratives as potential regulatory moat-seeking, and a gradual shift from outright non-shorting to limited partial hedges.[1][2]

Searches for January–June 2026 yielded limited specific public comments matching the criteria (new data post-April 3), suggesting his most prominent, detailed interventions clustered in July–October. His own podcast The Real Eisman Playbook (including Weekly Wrap episodes) and CNBC/Bloomberg-style interviews served as primary venues; no prominent social media posts surfaced in results. All quoted material below derives from direct interviews or his podcast episodes (or contemporaneous reporting thereof) and is treated as verified unless noted.[3]

July–August 2026: Achilles' Heel and Price War Concerns Emerge

Eisman highlighted the AI trade's heavy dependence on OpenAI and Anthropic as its core vulnerability ("Achilles' heel"), while expressing concern over cheaper Chinese/open-weight models potentially triggering price wars. He repeatedly stated he was not yet shorting the broader trade, citing lack of confirmatory data akin to pre-2008 subprime metrics, and described himself as "quite long" or not ready for a major short.[4][5]

• July 29, 2026: On The Real Eisman Playbook, Eisman said he would be "petrified" running OpenAI or Anthropic due to pricing pressure from models like Moonshot AI's Kimi K3 ($3 per million input tokens vs. higher OpenAI/Anthropic rates), calling it a signal of impending price war. He pressed guests on the lack of moats.[3]
• August 11, 2026: CNBC Fast Money — "The futures of these massive companies [hyperscalers], in a sense, are a bet that OpenAI, Anthropic are going to succeed." He warned of potential price wars from Chinese open-weight models taking share and noted the 70% AI revenue concentration (25–35% of cloud revenue) at Microsoft, Amazon, Alphabet, and Oracle.[5]
• Mid-August 2026 (podcast/Weekly Wrap referenced in reporting): Reiterated the 70% concentration figure and Achilles' heel; stated calls for a crash were "premature" due to no equivalent of Moody's data showing deterioration; remained "quite long."[4]
• Late August 2026 (interview with David Lin): Hypothetical — "If tomorrow OpenAI failed, the U.S. economy, I think, would go into an immediate recession and the market would have a massive correction." Emphasized it was not a prediction; described OpenAI as the "weak sister"; noted Anthropic's stronger growth trajectory at the time.[6]

Implication for market participants: Concentration risk was framed as the dominant near-term threat rather than valuations per se; early positioning favored monitoring usage/revenue trends at the two labs over aggressive bearish bets.

September 2026: Manufactured Crisis Narrative and Durability Questions

Eisman escalated criticism of AI safety/extinction warnings as a strategic ploy amid eroding moats ("token maxing is over," open-weight models gaining share). He affirmed real demand in infrastructure while questioning long-term durability due to potential price competition. His podcast featured dedicated episodes on these themes.[7][2]

• September 17, 2026: CNBC Squawk Box full interview — Called Terminator/extinction fears "nonsense" or "garbage." "I think these companies are very nervous. They realize that there are no moats around their business whatsoever, and they're trying to manufacture a crisis that will create regulation and that they think they can then manipulate to create the moats, to create the duopoly that they want." Suggested companies should "postpone your IPO" if truly concerned about pace/safety. Discussed trimming some AI exposure but not adding.[7][8]
• September 18, 2026: The Real Eisman Playbook Weekly Wrap ("Why Dario Amodei and Sam Altman Are Faking the AI Doomsday Crisis") — Expanded on the manufactured-crisis thesis tied to regulatory capture.[9]
• ~September 21–22, 2026: Podcast commentary (reported) — "70% of AI revenue of the hyperscalers is just from OpenAI and Anthropic" (~25–35% of their cloud revenue); entire ecosystem dependent on the two; OpenAI the weaker of the pair.[10]
• September 28, 2026: The Real Eisman Playbook Ep. 77 with Cisco IR head Sam Badri — Affirmed the boom is "for real" based on Cisco's 18% quarterly revenue growth and $9.3B fiscal 2026 AI infrastructure orders; however, "the question is how durable is it?" Warned of price wars with cheaper alternatives ("You don’t need Mercedes to do every single task").[2]

Implication: Demand-side evidence (e.g., hyperscaler capex and vendor growth) supported continuation in the near term, but competitive erosion and narrative skepticism introduced downside optionality without triggering full bearishness.

October 2026: Shift to Partial Hedging Amid Ongoing Concentration Worries

Eisman described beginning limited risk reduction via partial hedges while remaining mostly long; major short call still viewed as premature. He recapped prior themes in a high-profile podcast.[1]

• October 2, 2026: Prof G Markets interview (with Ed Elson) — Recapped CNBC clip on manufactured crisis and open-weight competition. Noted he had "shorted part" of four AI-related holdings "against the box" over the past month (partial hedges on long positions without outright sales; specific holdings unnamed; not a recommendation). "I have some shorts. I’m mostly long." "I took down some risk because I think like everybody else, I’m kind of nervous about the whole AI narrative." "If anything bad happens to one of those two companies within the next year, everybody’s in trouble." Major implosion call remains "premature." Also discussed Anthropic IPO/S-1 scrutiny and high yields as potential correction catalyst.[11][1]

Implication: Tactical de-risking via low-friction hedges signaled heightened caution without abandoning the long bias; concentration on two labs remained the central thesis.

Summary Table of Key Statements (Post-April 2026 Focus)

Date	Venue	Key Quote/Position	Shorting/Hedging Note	Source(s)
July 29, 2026	The Real Eisman Playbook (podcast)	"If I was the head of Anthropic or OpenAI, I’d be petrified" re: Chinese pricing pressure/price war risk.	Not shorting (consistent with later statements).	[web:46]
Aug 11, 2026	CNBC Fast Money	Hyperscalers' futures "are a bet that OpenAI, Anthropic are going to succeed"; Achilles' heel + Chinese models/price war risk; 70% concentration.	Not shorting the trade.	[web:52]
Mid-Aug 2026	Podcast/Weekly Wrap (reported)	70% concentration; crash calls "premature"; no Moody's-like data; "quite long."	Explicitly not shorting yet.	[web:50]
Late Aug 2026	Interview (David Lin)	Hypothetical OpenAI failure → "immediate recession"; OpenAI "weak sister."	Not predicting failure; not shorting.	[web:51]
Sep 17, 2026	CNBC Squawk Box	Terminator fears "garbage"/"nonsense"; companies "trying to manufacture a crisis" for regulation/moats/duopoly; "token maxing is over"; open-weight models gaining.	Stopped adding; trimmed some exposure.	[web:2], [web:35]
~Sep 18–22, 2026	The Real Eisman Playbook episodes	Faking doomsday crisis; 70% hyperscaler AI revenue from two labs; OpenAI weaker.	No major change noted.	[web:8], [web:11]
Sep 28, 2026	The Real Eisman Playbook Ep. 77 (Cisco guest)	AI boom "for real" (Cisco growth/orders); durability in question due to price competition.	N/A (focus on demand).	[web:4]
Oct 2, 2026	Prof G Markets (podcast)	Recapped manufactured crisis; concentration risk ("everybody’s in trouble" if one lab falters); high yields as correction risk.	Shift: partially hedged four AI positions "against the box" (past month); "some shorts. I’m mostly long"; major short "premature."	[web:5], [web:1]

No secondhand or unverified quotes flagged in primary reporting; all align with direct appearances or his podcast. Earlier 2026 data was sparse in results.[1]

For competitors or new entrants: Eisman's evolution underscores monitoring lab-level usage economics and Chinese/open-weight pricing as leading indicators; infrastructure demand can coexist with margin compression risks. Partial hedging ("against the box") offers a template for nuanced de-risking without full exits. Further primary transcripts from his Substack/premium episodes or additional CNBC appearances would strengthen granularity on exact portfolio moves.

The public record supports Eisman’s concentration thesis most strongly as a claim about contracted future cloud growth, not as a claim about current recognized cloud revenue. Microsoft and Oracle have disclosed enough for the backlog math to be checked. Amazon and Google have not. The “70% of AI revenue / 25–35% of cloud revenue” line is a sell-side estimate he adopted, not a number any of the four companies has reported.

What Eisman actually argued

On August 12–13, 2026, Steve Eisman said research reports put OpenAI and Anthropic at about 70% of AI-related revenue at Microsoft, Amazon, Google, and Oracle, and at 25% to 35% of those companies’ cloud revenue. He also said Oracle’s roughly $600 billion backlog was about half OpenAI. His conclusion: the hyperscalers’ futures are, in effect, a bet that two loss-making labs succeed. He later said he could not independently confirm the 70% figure, but that Oracle made it “sound right.” [1] [2] [3]

The same week, Ed Zitron published a parallel argument and cited sell-side work: UBS models putting OpenAI plus Anthropic at 27% of Google Cloud revenue in 2026 and more than 48% in 2027, and Barclays putting the pair at 13% of AWS this year, rising to 18% next. Aggregated, those desks were the source of the “over 70% of hyperscaler AI revenue” line. Zitron and Bloomberg also estimated that Microsoft’s disclosed OpenAI revenue was roughly 70% of Microsoft’s own AI business. [4] [5]

That distinction matters. “AI revenue” is not a GAAP line. “Cloud revenue” is. Mixing them is how a directional research estimate became a macro claim.

What the two labs have actually promised to buy

OpenAI’s named cloud purchase commitments add up to roughly $688 billion, none of it from an OpenAI filing:

• An incremental $250 billion of Azure, contracted in the October 2025 recapitalization that left Microsoft with about 27% of OpenAI Group, valued near $135 billion. Microsoft gave up its right of first refusal on new compute. [6]
• About $300 billion of Oracle capacity over roughly five years starting in 2027, tied to 4.5 gigawatts, as reported by the Wall Street Journal. Oracle has not isolated that contract in its own filings. [7] [8]
• $38 billion over seven years with AWS (November 2025), expanded in February 2026 by another $100 billion over eight years, alongside Amazon’s $50 billion investment ($15 billion upfront, $35 billion conditional). [9] [10]

Anthropic’s contractual floors with the three clouds are smaller than the headlines, and the gap between “expected spend” and “must pay” is the whole risk. A confidential IPO prospectus seen by Reuters puts infrastructure obligations at at least $518 billion over about a decade with six partners. About 80% is non-cancelable or payable regardless of usage. The cloud floors are: [11]

• Google: $111.1 billion, April 2026–July 2033. “If our actual spend falls short, we must pay Google the difference.”
• Amazon: $110 billion, May 2026–April 2036, same shortfall clause. This matches the April 2026 announcement of more than $100 billion over ten years and up to 5 gigawatts of Trainium, paired with a fresh $5 billion Amazon investment (cumulative investment then $13 billion, with up to $20 billion more available). [12] [13]
• Microsoft: $31.4 billion, November 2026–May 2033, cancelable only for Microsoft’s uncured material breach. That sits on top of the November 2025 announcement of $30 billion of Azure capacity.
• Broadcom-linked equipment leases: $161.2 billion, largely non-cancelable. That is not cloud revenue for Microsoft, Amazon, or Google.
• xAI: up to $84.5 billion, mostly cancelable on 90 days’ notice — the exception that shows how hard the other contracts are.

In May 2026, The Information reported a separate $200 billion Anthropic commitment to Google Cloud over five years, more than 40% of the cloud backlog Alphabet had just disclosed (then a bit over $460 billion). Reuters carried that report. The prospectus minimum of $111.1 billion is a floor, not a contradiction of a larger expected-spend figure, but they should not be added together. [14] [15]

Against that bill, Anthropic’s 2025 results in the same filing are small: revenue nearly $4.6 billion (up 12-fold), operating losses above $8 billion. Consumption-based Claude usage was about $3.8 billion; subscriptions were $789 million. [16]

Backlog is where the dependence is measurable

By August 2026 the four providers’ contracted backlogs were, on Cloud Wars’ compilation of company figures: Microsoft commercial remaining performance obligations $678 billion, Oracle RPO $638 billion, Google Cloud $514 billion, AWS $496 billion — about $2.3 trillion combined. [17] Microsoft’s investor metrics confirm the $678 billion year-end figure and Microsoft Cloud revenue of $214.4 billion in fiscal 2026, with Azure and other cloud services up 41%. [18]

Microsoft is the only company that has quantified a lab’s share of backlog.

• At December 31, 2025, commercial RPO was $625 billion. CFO Amy Hood said about 45% was OpenAI — roughly $281 billion — after the $250 billion Azure commitment landed. The other 55%, about $350 billion, grew 28%. [19] [20]
• At June 30, 2026, RPO was $678 billion, up 84%. Hood said it increased 25% excluding OpenAI. Starting from the prior-year $368 billion, that implies roughly $460 billion ex-OpenAI and about $218 billion, or ~32%, still tied to OpenAI. That 32% is arithmetic from her comment, not a second direct disclosure. Either way, most of the year’s backlog growth came from one customer. [21] [22]

The same 10-K booked $24.1 billion of fiscal 2026 revenue from commercial arrangements with OpenAI, including revenue-sharing payments, and $6.0 billion of receivables from OpenAI at year-end. Microsoft had funded $11.9 billion of a $13 billion commitment. OpenAI was about 7% of Microsoft’s $331.8 billion of total revenue. Set against Azure’s first $100 billion year, the $24.1 billion is about a quarter of Azure only if all of it is compute — and the filing says it is not. [22] [23]

Oracle is the cleanest credit case, and it matches Eisman’s “half the backlog” line.

• RPO went from about $138 billion to $455 billion in one quarter in 2025, then $523 billion, $638 billion at May 31, 2026 (up 363%), and $664 billion by August 31, 2026. [24] [25]
• On July 9, 2026, S&P cut Oracle to BBB-, one notch above junk, and called OpenAI a key credit risk. S&P said OpenAI was roughly half of the $638 billion RPO. If OpenAI cannot pay, Oracle is left with data-center leases it may be unable to exit or must re-lease on worse terms. Bank of America analysts have made the same “more than half” estimate. [26] [27] [28]
• Recognized revenue tells the opposite story. Oracle has disclosed that no customer was 10% or more of total revenue in fiscal 2024, 2025, or 2026. The OpenAI contract is reported to start in 2027. Concentration is in the order book, not the income statement. [29]

The duration mismatch is the mechanism. Oracle signs 15- to 19-year building leases and long power contracts, buys the chips, and sells a roughly five-year contract to an unrated customer. Uncommenced leases were reported around $248–288 billion. Fiscal 2026 free cash flow was about negative $23.7 billion after $55.7 billion of capex. S&P projected a fiscal 2027 free-operating-cash-flow deficit near $42 billion. [30] [26]

Amazon and Google have not allocated backlog by customer. Amazon’s performance obligations, primarily AWS, reached $496 billion by June 2026. The filing identifies the $100 billion OpenAI expansion and the greater-than-$100 billion Anthropic expansion but does not say how much of the $496 billion they represent. If both are fully inside that balance, the two labs would be on the order of half of AWS’s backlog — an inference, not a disclosure. [31] [32] Google Cloud’s backlog was a bit over $460 billion in the quarter The Information used for the “more than 40%” Anthropic claim, and management later cited $514 billion. [33] [14]

A May 2026 Information analysis, carried by Reuters, said contracts involving the two labs already accounted for more than half of about $2 trillion of major-cloud backlog. Later backlog growth would dilute that share unless new lab contracts were added at the same pace. Directionally, a stack of OpenAI’s ~$688 billion of named cloud commitments plus Anthropic’s ~$253 billion of cloud floors (or ~$340 billion if the $200 billion Google figure is used instead of the $111 billion floor) is on the order of $0.9–1.0 trillion, or roughly 40–50% of a $2.0–2.3 trillion combined book. That is the quantitative core of Eisman’s argument, and it is about signed future revenue, not last year’s cloud P&L. [14]

How well the 70% and 25–35% claims hold up

Claim	What public data can actually show	Verdict
70% of the four companies’ AI revenue	No company reports “AI revenue.” Bloomberg’s ~70% is a Microsoft-only estimate: $24.1 billion from OpenAI versus an extrapolated ~$34 billion AI business. UBS/Barclays figures cited by Zitron are models.	Directional for Microsoft’s labeled AI business; not audited across four firms
25–35% of current cloud revenue	Microsoft: $24.1 billion is ~11% of $214 billion Microsoft Cloud, and ~24% of a $100 billion Azure year only if revenue-share is ignored. Oracle: no 10% customer in recognized revenue; the big contract starts in 2027. AWS/GCP: undisclosed. Barclays’ 13–18% of AWS and UBS’s 27% of Google Cloud in 2026 sit at or below the range.	Not demonstrated in current revenue. Plausible as a forward blend once take-or-pay ramps
Half of Oracle’s backlog is OpenAI	S&P: roughly half of $638 billion. WSJ: ~$300 billion contract.	Supported
OpenAI was 45% of Microsoft’s commercial RPO	Hood, December 2025 quarter, on $625 billion	Supported, as of that quarter; implied share later fell toward ~32% as other RPO grew
Two labs are ~half of combined hyperscaler backlog	Information/Reuters, May 2026, on a ~$2 trillion book. Later books are larger; customer splits at AWS and Google remain undisclosed	Supported as an estimate of contracted growth, not as a current-revenue fact

Three accounting traps make the headline percentages easy to overread.

First, Microsoft’s $24.1 billion includes revenue share. Microsoft earns when OpenAI sells to end customers, not only when OpenAI rents GPUs. Treating that line as pure cloud consumption overstates Azure dependence and understates how much of the “AI revenue” is a royalty on the lab’s own sales. [22]

Second, the labs are also distribution channels. In 2025, Anthropic routed $2.16 billion, or 47% of revenue, through Amazon and Google marketplaces, up from 11% in 2023 and 32% in 2024, and paid about $351 million back in distribution fees — roughly 16 cents per marketplace dollar. Those partners collected 60% of $909 million of customer bills outstanding. Two unnamed customers were each 12% of Anthropic’s revenue. Some “hyperscaler AI revenue” is a toll on the lab’s end customers, who are the real demand. Counting lab spend and enterprise AI spend as separate piles double-counts. [16]

Third, backlog is not cash. Oracle expects only about 12–13% of its RPO to convert within twelve months. Microsoft’s RPO duration has been about 2–2.5 years. A five-year OpenAI contract that starts in 2027 can re-rate a stock today and still not be revenue for several quarters. [28] [34]

The claim that holds is narrower and more serious than the soundbite: the incremental cloud growth that re-rated these stocks is unusually dependent on two customers. The claim that does not hold, on disclosed revenue, is that those two already are 25–35% of cloud sales at all four companies.

Circular financing and counterparty risk

The loop is the same at each node. A supplier invests equity (or hands over warrants). The lab commits to spend a multiple of that equity back on the supplier’s cloud or chips. The supplier books revenue, and often a mark-to-market gain on the equity, while the lab funds the spend with the next round. Bloomberg has mapped the pattern across Microsoft–OpenAI, Amazon/Google–Anthropic, Nvidia, AMD, and Oracle/Stargate. [35]

Specific loops that are on the record:

• Microsoft–OpenAI. More than $13 billion invested; 27% stake marked near $135 billion; $250 billion incremental Azure commitment; $24.1 billion of fiscal 2026 related-party revenue. The cash Microsoft put in is a fraction of the revenue and equity value it has booked back. [6] [22]
• Amazon–Anthropic. Ars Technica, the Financial Times, and TechCrunch all described the April 2026 deal as circular: $5 billion in (on top of $8 billion already invested) against more than $100 billion of AWS spend, much of it on Amazon’s own Trainium chips. By June 30, 2026, Amazon carried the Anthropic stake at about $190 billion, and one period’s “other income” of $53.4 billion was primarily Anthropic valuation marks. Cloud revenue and investment gains move together. [36] [37] [38] [31]
• Amazon–OpenAI. $50 billion investment, only $15 billion upfront, against a $100 billion expansion of an existing $38 billion cloud deal, including 2 gigawatts on Trainium. William Blair estimated an even spend would be about $17 billion a year, roughly 11% of expected 2026 AWS revenue — a single-customer growth slug funded in part by the vendor. [10] [39]
• Nvidia and AMD. Nvidia put $30 billion into OpenAI’s February 2026 round after a discussed commitment of up to $100 billion was scaled back. AMD agreed to buy up to $5 billion of Anthropic stock alongside more than $20 billion of expected compute supply. The chipmaker’s largest customers are also its investees. [39] [11]
• Alphabet–Anthropic. Up to $40 billion of equity, reported alongside the multi-gigawatt TPU deal, against a take-or-pay cloud floor of $111.1 billion and a reported $200 billion expected spend. [14]

Anthropic’s own prospectus states the conflict directly: Amazon, Google, and Microsoft are simultaneously investors, customers, cloud providers, distributors, and competitors, with incentives that “may not be fully aligned.” [11]

The counterparty risks analysts and rating agencies have actually named:

1. Ability to pay, not willingness to cancel. Google and Amazon minimums are take-or-pay. Microsoft’s Anthropic contract is non-cancelable except for Microsoft’s breach. The protection for the hyperscaler is only as good as the lab’s balance sheet. Anthropic’s 2025 operating loss exceeded $8 billion on $4.6 billion of revenue, against a $518 billion buildout. OpenAI’s named cloud commitments of ~$688 billion similarly dwarf any reported revenue run-rate. Both labs have to keep raising capital, or go public, for the RPO to become cash. [11] [16]

2. Oracle is the first loss-absorber. S&P’s failure path is specific: unpaid OpenAI contracts leave Oracle holding long leases and power deals it cannot easily exit. Credit markets have already priced some of that. Five-year CDS have traded at stressed levels, 2056 bonds have yielded above 8%, and debt tied to the $18 billion Project Jupiter campus has traded around 89–90 cents on the dollar. Some banks reportedly declined Stargate-linked financings where Oracle was the anchor tenant, citing concentration. [26] [40] [28]

3. Correlated marks. A price war — Eisman’s China open-weight scenario — would hit lab equity values (Amazon’s $190 billion Anthropic carrying value, Microsoft’s $135 billion OpenAI stake, Alphabet’s Anthropic marks), the collectibility of cloud RPO, and Nvidia’s order book in the same quarter. The equity upside and the cloud receivable are not independent hedges. They are the same bet, booked twice.

4. Stranded capacity is not equal across the four. S&P’s point, which Eisman has echoed, is that AWS, Google, and Microsoft have internal workloads (Search, YouTube, Gemini, Copilot, ads, retail) that can absorb unused GPUs. Oracle does not, at anything like the same scale. Microsoft’s own disclosure cuts both ways: ex-OpenAI RPO still grew 25–28%, and Azure has been capacity-constrained, so the franchise is not only two customers. But the acceleration is. [27] [20]

5. Off-balance-sheet temptation. Eisman has argued that special-purpose vehicles, guarantees, and lease structures are being used to keep AI project debt away from ratings — the incentive S&P’s July downgrade made explicit for Oracle. The economic exposure remains even when the debt does not consolidate. [41]

What this means if you are underwriting the trade

Treat backlog share and revenue share as different securities. Oracle’s equity and credit are a direct underwriting of OpenAI’s ability to fund a contract that starts in 2027 and is about half the order book, against leases that run into the 2040s. Microsoft’s risk is large but diluted: OpenAI was 45% of commercial RPO at the peak disclosure and still the majority of backlog growth, yet it was only about 7% of total company revenue, and part of the $24.1 billion is a royalty, not a GPU rental. Amazon and Google have the same circular structure, but investors cannot see the customer split; the binding evidence is the labs’ own take-or-pay floors, not the hyperscalers’ 10-Qs.

The competitive implication is uncomfortable for anyone selling cloud against these four. The incremental AI capacity is not being won in an open market. It is being pre-sold to two customers who are also financed by the sellers, on contracts that force payment even if usage misses. A third-party cloud, or a neocloud, is not competing only on price and chips. It is competing with a vendor-financed order book that already accounts for something like half of the industry’s contracted growth. The hedge inside the hyperscalers is their non-AI franchise and internal demand. The thing that is not hedged is the re-rating: if either lab’s funding window closes, the backlog that justified the capex does not disappear from the slide. It becomes a credit loss and a pile of unfilled buildings, first and hardest at Oracle.

---

Recent Findings Supplement (October 2026)

Steve Eisman has reiterated and partially hedged his concentration thesis in September–October 2026 commentary, estimating that OpenAI and Anthropic together drive roughly 70% of AI-related revenue at Microsoft, Amazon, Google, and Oracle (equating to 25–35% of those firms’ total cloud revenue).[1]

This builds directly on his August 2026 statements and is supported by newly disclosed contract sizes and backlog attributions rather than broad assertions. Eisman has begun reducing exposure by shorting portions of AI holdings “against the box” while watching for price wars from open-weight models and signs that one or both labs could falter.[1]

• Anthropic’s September 2026 confidential IPO prospectus (seen by Reuters) details $518 billion in planned decade-long infrastructure spend, with ~80% non-cancellable or payable regardless of usage.[2]
• Microsoft’s FY2026 commercial remaining performance obligations (RPO) reached $678 billion as of June 30, 2026 (up 84% YoY); excluding OpenAI, growth was only 25%, implying OpenAI accounts for a large share of incremental backlog (inferred ~$217 billion+ tied to the lab).[3]
• S&P estimates roughly half of Oracle’s $638 billion backlog comes from OpenAI.[4]

This data bolsters Eisman’s core claim of heavy dependence. Public filings and reporting now quantify the scale of take-or-pay-style commitments that tie hyperscaler growth directly to the labs’ continued spending power. For competitors or entrants, the mechanism is clear: winning meaningful share requires either displacing these locked-in contracts or capturing diversified non-lab demand that is growing more slowly.

Anthropic’s September 29, 2026, IPO prospectus reveals $518 billion in AI infrastructure obligations over the next decade, including $111.1 billion with Google (April 2026–July 2033), $110 billion with Amazon (May 2026–April 2036), and $31.4 billion with Microsoft (November 2026–May 2033, non-cancellable except for material breach), plus $161.2 billion in largely non-cancellable Broadcom equipment leases.[5]

Roughly 80% of the total is structured as minimum-spend or pay-regardless-of-usage terms, giving cloud providers high visibility into future revenue but exposing them to Anthropic’s ability to fund the obligations. The company reported ~$4.6 billion in 2025 revenue against >$8 billion operating losses.[1]

• Additional details include AMD committing up to $5 billion in stock purchases and supplying >$20 billion in capacity.[2]
• Earlier 2026 expansions (e.g., >$100 billion AWS deal announced April 2026 for up to 5 GW) are folded into these aggregates.[6]

The mechanism is take-or-pay contracting that converts lab funding needs into guaranteed hyperscaler revenue streams. Implication: Any slowdown in Anthropic’s fundraising or model monetization directly pressures the clouds’ realized returns on capex. New entrants face a high bar—labs are locking in multi-year capacity with penalties for under-utilization, leaving less flexible demand for alternative providers in the near term.

OpenAI’s post-restructuring commitments include an incremental $250 billion Azure purchase obligation tied to Microsoft’s ~27% stake in the October 2025 recapitalization, a reported ~$300 billion five-year Oracle cloud deal (starting 2027, 4.5 GW), and an AWS expansion adding $100 billion over eight years to a prior $38 billion commitment.[7]

Microsoft booked $24.1 billion in revenue from OpenAI arrangements in FY2026 and reported OpenAI owing it $6 billion as of June 30; OpenAI-related deals represent a substantial portion of Azure growth and ~45%+ of Microsoft’s commercial RPO in recent periods.[3]

• Broader OpenAI compute commitments have been cited in the $600 billion–$1.4 trillion range through 2030 (with revisions noted in reporting).[8]
• Microsoft’s Azure crossed $100 billion annual revenue run-rate in FY2026, with OpenAI arrangements equating to roughly a quarter of that in some analyses.[3]

Public data therefore shows OpenAI as Microsoft’s largest single AI customer and a material driver of backlog growth across multiple providers. The circular element is explicit: Microsoft’s equity stake and financing coincide with OpenAI’s spend commitment back to Azure. For market participants, this creates concentrated counterparty exposure—OpenAI’s cash burn ($3.7 billion in Q1 2026 alone; projected heavy losses through 2030) means its ability to honor obligations depends on ongoing external capital raises.[3]

Hyperscaler backlogs and revenue forecasts now embed large contributions from the two labs, with UBS estimating OpenAI + Anthropic could account for 48% of Google Cloud’s 2027 revenue (~$84–100 billion), plus ~$40 billion at AWS and at least $50 billion at Microsoft Azure.[9]

Goldman Sachs (September 25, 2026 note) calculates the five largest U.S. hyperscalers need ~$300 billion in annual AI revenue to break even on projected 2026 AI infrastructure spend of ~$800 billion, against only ~$70 billion in current AI cloud revenue above pre-buildout levels—a ~$230 billion annual shortfall.[10]

• Oracle has raised significant debt/leases amid its OpenAI exposure; S&P links half its backlog to the lab.[4]
• Eisman and others note Nvidia’s own concentration (70% of revenue from five hyperscaler customers whose AI revenue loops back to the labs).[9]

The supporting mechanism is that capex and backlog growth are front-loaded on lab commitments whose realization depends on the labs’ solvency and usage. This strengthens Eisman’s argument with quantitative backing from filings and analyst models. For entrants or competitors, the implication is that near-term cloud/AI infrastructure demand is heavily pre-committed and concentrated; diversified or lower-cost alternatives must overcome both contractual lock-ins and the labs’ preference for established hyperscalers with financing ties.

Analysts and the Bank for International Settlements (BIS Bulletin 137, October 1, 2026) have highlighted circular-financing and counterparty risks, noting that >55% of AI funding (2021–2025) came from other AI firms, with nearly half of intra-AI deal value tied to commercial supply relationships.[11]

Examples include Microsoft’s equity in OpenAI paired with OpenAI’s $250 billion Azure commitment; Amazon/Google stakes in Anthropic alongside their large compute contracts; and chipmaker investments (Nvidia, AMD, Broadcom) that loop back as demand for their products. BIS warns these self-referential structures make reported demand partly endogenous and can amplify macroeconomic and contagion risks in stress scenarios.[12]

• Eisman has flagged Enron-like off-balance-sheet techniques and guarantees resurfacing in AI financing (September 2026).[13]
• Labs’ losses and funding dependence create direct credit exposure for hyperscalers holding the corresponding receivables and RPOs.

The risk mechanism is interconnected balance sheets where one party’s distress (lab funding shortfall or model commoditization) simultaneously hits revenue, equity values, and debt service across the chain. Public data supports Eisman’s warning: the concentration is real and quantifiable, and the circular elements introduce fragility not present in traditional enterprise cloud growth. New participants must navigate these closed loops—either by securing independent funding/customer bases or by offering materially cheaper alternatives that can pry open locked capacity. Overall, recent disclosures (especially Anthropic’s prospectus and backlog details) provide concrete numerical support for the dependence thesis while underscoring the elevated counterparty and circular risks analysts have flagged.

Steve Eisman’s 2026 AI view is not “the boom is fake.” It is that model companies have no pricing moat, a price war is already visible in API lists, and the public AI complex is a credit bet on two private labs. Infrastructure can still get paid while that bet holds. He has not made the outright short on the trade.

Model companies have no durable pricing moat

Eisman’s mechanism is switching costs near zero. Users move among ChatGPT, Gemini, and Claude, so the lab with the best model this quarter does not own next quarter’s price. On CNBC on September 17 he said token-maxing is over, open-weight models are taking share, and “there are no moats around their business whatsoever.” On The Real Eisman Playbook he put it more bluntly: “There are no pricing moats in this business. Today I have the best LLM and tomorrow yours is better and cheaper.” In a June Fortune interview he extended the same logic to hyperscalers selling AI services: even if AI is “the greatest thing since the invention of the printing press,” providers have nothing to shield them from price competition. [1] [2] [3]

He treats the safety slowdown as a substitute moat, not a forecast. After Dario Amodei and Sam Altman called for slower capability gains, Eisman called the “Terminator” framing “garbage” and said the labs are trying to manufacture a crisis that produces regulation they can shape into a duopoly. The tell, in his view, is that neither side is postponing IPOs or capital raises. He later conceded he does not know their motives. A secondary reading from his commentary is that Nvidia’s interest in the open-weight ecosystem (he floated a Hugging Face deal as an “insurance policy”) is a hedge against a lawsuit or a closed-lab failure, not evidence of a model moat. [4] [5]

What this means for anyone competing: at the model layer, Eisman is betting capability converges and price does the work. Distribution, enterprise contracts, and regulated use cases can still matter, but he does not treat them as castle walls.

The price-war call is conditional, not a crash call

The prediction has two speeds. A price war among frontier and open-weight models is, in his words, already showing “signs.” A price war severe enough to break the infrastructure boom has not arrived.

In late July he pointed at Moonshot’s Kimi K3 at $3 per million input tokens versus $5 for OpenAI’s GPT-5.6 Sol and $10 for Anthropic’s Claude Fable 5, plus open weights that let customers leave the platform. “If I was the head of Anthropic or OpenAI, I’d be petrified. That spells to me price war.” By late September, after Cisco’s AI orders accelerated, he agreed demand is real—“you don’t need Mercedes to do every single task”—but said a full-blown war with Chinese and open-weight models is what would change the story. “If that were to happen, maybe the story would change. But until then, the story probably continues.” On October 2 he repeated that calling an implosion is still premature. [6] [7] [8]

The data matches the first half of that call and complicates the second.

• July 30: OpenAI cut GPT-5.6 Luna by 80%, to $0.20 / $1.20 per million input/output tokens; Terra fell 20% to $2 / $12. CFO Sarah Friar later said the Luna cut helped drive roughly a tenfold usage increase. [9] [7]
• August 10: Anthropic locked Claude Sonnet 5 at $2 / $10 and canceled a scheduled rise to $3 / $15. [10]
• September 1: Anthropic cut Fable 5.1 cache-read pricing 75%, from $1.00 to $0.25 per million tokens, aimed at agent workloads. [10]
• September 22: OpenAI launched GPT-6 Sol at $2 / $10 and GPT-6 Luna at $0.10 / $0.50, about half prior-generation rates. Anthropic launched Claude Opus 5.5 at $4 / $20, 20% below Opus 5, and said effective run-cost was about 40% lower. Flagship tiers (GPT-6 Astra, Claude Fable-class) stayed near $10 / $50. [11] [12] [13]
• Chinese open and API prices remain the floor. Kimi K3 lists at $3 input / $15 output, with cached input at $0.30. DeepSeek’s flash tier has been cited near $0.14 / $0.28. [14] [9]

Inference costs at a fixed capability level are falling faster than the list-price headlines. Epoch AI’s September 22, 2026 analysis finds the price of a given performance level down about 47% per quarter since 2023, or roughly 13 times per year—faster than DNA sequencing, compute, batteries, or historical electricity. One concrete path: o3 scored 75% on GPQA Diamond in early 2025 at about $0.30 per question; a GPT-5.6-class model matched it under 18 months later at $0.0004, a 725-fold drop. Stanford’s earlier series showed GPT-3.5-class inference falling from $20 per million tokens in November 2022 to $0.07 by October 2024. A usage-weighted Silicon Data index hit $0.97 per million tokens on September 1, 2026. [15] [16] [17]

The non-obvious implication is the one Eisman half-concedes: cheaper tokens can raise consumption. Friar’s tenfold Luna usage jump is the Jevons case. If agent workloads burn more tokens as unit prices fall, hyperscaler and Nvidia revenue can rise even as lab gross margins compress. That is why he can say the boom is “for real” and still say model companies have no moat. The test that has not cleared is whether open-weight share gets large enough in dollars, not benchmarks, to force the labs to break their take-or-pay cloud contracts. One September estimate put open models at roughly 6–11% of global AI revenue versus a much larger closed-model pool. That is pressure, not displacement. Treat that share figure as an analyst reconstruction, not a filing. [18]

The chain is a bet on two private balance sheets

Eisman’s load-bearing claim is concentration, not valuation. He says about 70% of AI-related revenue at Microsoft, Amazon, Alphabet, and Oracle comes from OpenAI and Anthropic, equal to roughly 25–35% of those companies’ cloud revenue, and that something like half of Oracle’s contracted backlog is OpenAI. On August 12 he attributed the 70% figure to research-firm estimates, not to a disclosure he had seen. On October 2: “If anything bad happens to one of those two companies within the next year, everybody’s in trouble.” He calls OpenAI the weaker of the two and, in late August, said an OpenAI failure would put the U.S. economy into an “immediate recession.” He is explicit that five years out the customer base should be more diversified, so the window that matters is the next year or so. [8] [19] [20]

Public filings support concentration. They do not confirm a single 70% number across four clouds.

• Microsoft’s fiscal 2026 filing recorded $24.1 billion of revenue from commercial arrangements with OpenAI, including revenue-sharing, about 7% of Microsoft’s $331.8 billion total revenue. Azure topped $100 billion. The $24.1 billion is not pure compute, so it cannot be read as “70% of Azure.” Commercial remaining performance obligations were $678 billion, up 84%; excluding OpenAI, RPO grew 25%. [21]
• A Reuters review of Anthropic’s confidential prospectus: 2025 revenue nearly $4.6 billion, about 12 times 2024; operating loss more than $8 billion; compute and infrastructure spend $7.33 billion; cash and short-term investments $20.28 billion at year-end 2025. Net loss near $42 billion included roughly $34 billion of non-cash charges tied to financing that could convert into shares. Future cloud, compute, and infrastructure obligations: at least $518 billion, about 80% non-cancelable or payable regardless of use—including at least $111.1 billion with Google, $110 billion with Amazon, $31.4 billion with Microsoft, and about $161.2 billion of largely non-cancelable Broadcom-related equipment leases. Forty-seven percent of 2025 sales were routed through Amazon and Google. Two unnamed customers each accounted for 12% of Anthropic’s own revenue, and the company warned many large clients are not locked into long-term contracts. [22] [23] [24]
• Separate analyst work (UBS, Barclays, and Ed Zitron’s compilation) puts the two labs at a very large share of hyperscaler AI revenue, with Microsoft’s OpenAI dependence the clearest disclosed case and Google/AWS shares estimated rather than filed. Zitron’s own addition across Microsoft, Google, Amazon, Oracle, and SpaceX landed nearer 64% of estimated AI revenue, not a uniform 70%. [25] [26]
• Oracle: reported accounts describe an OpenAI compute commitment on the order of $300 billion over five years against roughly $638 billion of contracted future business, which is the factual core of Eisman’s “half the backlog” line. Oracle raised $43 billion of debt in fiscal 2026 while free cash flow fell to negative $23.7 billion. S&P cut the rating to BBB− on July 9. Debt tied to the $18 billion Project Jupiter data center has traded around 90 cents on the dollar. [20] [27]

The asymmetry he is pointing at is contractual. Anthropic’s spend is largely take-or-pay. Its customers, by its own risk factors, often are not. If token prices keep falling and usage does not fill the minimums, the labs pay the difference. That is a credit problem for Oracle and the neoclouds before it is a multiple problem for Nvidia.

He separates balance-sheet stress inside the public complex. On October 2 he said Oracle is the weakest among the large names, not necessarily weaker than neoclouds, and that he does not lose sleep over Meta, Google, or Microsoft at current ratings. The financing technique that bothers him is off-balance-sheet project debt: Meta owns 20% of the Hyperion venture, Blue Owl-managed funds own 80%, and a Blue Owl vehicle raised $27.3 billion of debt while Meta keeps lease and overrun exposure. Ernst & Young flagged the accounting as a critical audit matter. Eisman called the pattern “reminiscent of bad times past” without accusing fraud. [28] [27]

Infrastructure is his preferred seat only while the labs keep spending. In June he favored Nvidia, Arista, and Cisco—chips and networking that must be bought no matter which chatbot wins—over Meta, Oracle, Microsoft, and Alphabet as AI service vendors. By late September, Cisco’s quarter (revenue up 18% to $17.3 billion, $4 billion of hyperscaler AI infrastructure orders, $9.3 billion fiscal 2026 AI orders) was his evidence that the boom is real. The durability question is still the labs. [3] [7]

Every named long, short, hedge, and concern in 2026

Eisman is a commentator and personal investor, not a current 13F filer. Searches of 13F trackers turn up historical roles at FrontPoint and Emrys Partners, not a 2026 institutional portfolio. Everything below is self-disclosed on CNBC or his podcast. The four AI hedges are unnamed on purpose. Against-the-box shorts would not show up as reduced longs even if he filed.

Name	Stance in 2026	Reason he gave
Fair Isaac (FICO)	Explicit short, disclosed April 30 on CNBC; still the short he discussed October 2	Raised mortgage-score prices aggressively (he has said ~500% over many years, and on October 2 “1,600%” over five years), “ticked off” the lending ecosystem and FHFA’s Bill Pulte. Fannie, Freddie, and FHA can use VantageScore. He cited a cost gap on the order of ~$2,000 per 100 applications versus about $99, and later Rocket Mortgage moving to VantageScore. Not an AI short. “Don’t piss off your regulator.” [29] [28]
Alphabet (GOOGL)	Sold a long-held stake “a couple of months” before July 27; sat in cash; not described as a short	Cut AI exposure. “It’s all one trade.” More than half of a typical 60/40 equity sleeve is tech/AI, and much new bond issuance is AI-related. Defensive rotation into something like Clorox does not work because buyers want AI or nothing. Came as Alphabet lifted 2026 capex toward $180–205 billion. [30] [31]
Nvidia (NVDA)	Long as of late August, “less long,” not short. Likely inside the unnamed AI book he has since partially hedged	Picks and shovels. Still necessary whichever lab wins. He has also warned that an OpenAI or Anthropic problem would hit Nvidia with the rest of the chain. In October he dismissed Michael Burry’s depreciation-schedule critique as “too academic” relative to lab failure. [20] [32]
Four unnamed AI holdings	Partial short against the box over the month before October 2	Hedge, not a thesis that the trade implodes. He said he is mostly long, nervous, and unwilling to make the big short. Tax constraint he stated himself: unwind by the end of January or the IRS can treat it as a sale. He did not recommend the tactic. [8] [28]
Cisco (CSCO)	Preferred infrastructure name; hosted Cisco IR; not a disclosed personal position in the sources reviewed	Evidence the boom is real: growth re-accelerated from ~5% doubts to 18%, with a visible hyperscaler AI order book. Durability still depends on whether cheaper models kill token demand or just mix. [7]
Arista (ANET)	Preferred with Nvidia and Cisco in the June Fortune framing; not a confirmed 2026 holding	Same picks-and-shovels logic: networking gear hyperscalers must buy regardless of which model wins. [3]
Anthropic	Would not invest in the IPO at the valuation being discussed (~$2 trillion in market talk); concern, not a shortable stock	No moat, huge non-cancelable compute bills, losses, and a safety narrative he reads as regulatory capture. He wants the S-1, which is still confidential. [8] [4]
OpenAI	Concern; “the weaker company”; not investable as a public stock in 2026	Concentration risk for the whole chain, postponed IPO, price-war exposure, dependence on continued capital raises. August: failure would mean immediate recession and a massive correction. [1] [20]
Oracle (ORCL)	Concern; weakest large balance sheet in the AI trade	BBB−, negative free cash flow, large OpenAI-linked backlog, off-balance-sheet project debt, rating defense. He does not claim it is weaker than neoclouds. [28] [27]
Meta (META)	Financing concern, not a disclosed short	Hyperion / Blue Owl structure keeps project debt off the balance sheet while Meta retains economic exposure. He groups Meta with hyperscalers that lack a service-layer moat. [27] [3]
Microsoft, Amazon	Structural concern via customer concentration; he has not disclosed a 2026 position	Same 70% / two-customer thesis. He has said he does not lose sleep over Microsoft’s or Google’s ratings the way he does Oracle’s. [28]
SpaceX	Valuation concern ahead of/around the IPO; not a disclosed short	June: expected valuation near 100 times revenue is “kind of crazy”; revenues comparable, in his joke, to Kellogg’s; likely to become a retail cult stock. [3]
Private equity and private credit software book	Sector concern, October 2	AI is eroding software moats (“SAASpocalypse”), buyout valuations already cut 50% or more, and a 2027 refinancing wall can wipe equity. This is the application-layer version of his model-moat argument. [33]
The market / 60-40 portfolios	Macro concern, not a net short	“It’s all one trade.” If the 10-year stays above 5%, a correction is “probably imminent.” He is not willing to short the AI complex outright. [30] [34]
P&C insurers / brokers	Discussed as a possible AI hedge via a guest (Ryan Tunis, September 7); not his disclosed book	Brokers face AI disintermediation risk; some carriers are self-help or cycle stories. Do not treat guest names (Travelers, Chubb, AIG, Kinsale, Trupanion) as Eisman positions. [35]

He also said in August he was not shorting the AI trade because he wanted lab financials after they go public. The October hedge is a partial de-risking of longs he already owns, which is consistent with that earlier line rather than a reversal into a Burry-style short. [36]

Which claims are testable, and what the data shows

Price war at the API layer — supported. Same-week cuts by OpenAI and Anthropic in September, an 80% Luna cut in July that raised usage about tenfold, permanent mid-tier pricing at $2 / $10, and Chinese open-weight prices below U.S. frontier lists are public. Frontier flagships have not been given away; the war is in the mid and cheap tiers, plus cache pricing for agents. That is exactly the “you don’t need Mercedes for every task” mechanism.

Inference-cost collapse — supported, and faster than list prices. Epoch’s ~13× per year at constant capability, the 725-fold GPQA example, and sub-$1 usage-weighted token indexes are independent of Eisman. Nvidia has also claimed up to 10× lower inference token cost on Rubin versus Blackwell from the second half of 2026. Falling cost is not the same as falling industry revenue.

“Token maxing is over” and open-weight models have taken big revenue share — not confirmed. Usage can rise when prices fall. Open-model revenue share, on the one reconstruction available, is still a small slice of dollars. Eisman may be early on share and right on the direction of pricing power. This is the claim most in need of a clean usage-versus-revenue series over the next two quarters.

No durable moat at the model layer — directionally supported, not settled. Repeated price cuts and open weights are evidence against pricing power. They are not evidence against distribution, enterprise lock-in, or a future regulatory moat. Anthropic’s own filing says customers can leave, which helps his case, while its growth (12× revenue in 2025, and later run-rate reports far above that trailing year) shows the franchise is not yet being competed away.

70% of hyperscaler AI revenue from two labs — plausible, not filed. Microsoft’s $24.1 billion OpenAI line and Anthropic’s take-or-pay cloud contracts make the dependency real. The round 70% across Microsoft, Amazon, Google, and Oracle is an estimate Eisman attributes to sell-side work. Independent compilations land in a wide band, roughly 50–70% of AI revenue depending on what is counted as “AI revenue.” Oracle’s OpenAI share of backlog is the cleanest single-customer fact in the set, and it is large.

OpenAI is the weaker lab — not yet testable in public numbers. OpenAI has not produced an equivalent prospectus. The observable facts that fit his ranking are the postponed 2026 IPO and Anthropic’s decision to keep filing. Revenue-run-rate headlines have at times shown Anthropic ahead, but those are press figures, not audited comparisons.

Infrastructure boom continues until the price war breaks lab spending — still intact as of October 3, 2026. Cisco’s order book and Nvidia’s data-center prints have not rolled over. Eisman’s own standard is lab health over the next year, not this quarter’s multiples. Polymarket, for what it is worth, has priced a multi-condition AI downturn by year-end well below 10% in the articles that cited it. That is sentiment, not evidence.

Anthropic can service $518 billion of obligations — the central unresolved credit test. Trailing revenue of $4.6 billion and $20 billion of cash against $518 billion of mostly non-cancelable commitments only works if the run rate keeps compounding and usage fills minimums. The prospectus reportedly points at very large 2028 revenue. That projection is the whole underwriting. If it misses, Oracle, Broadcom-linked leases, and the hyperscaler RPO lines are the transmission mechanism Eisman described. If it hits, his moat critique can be right and his “everybody’s in trouble” scenario can still be wrong, because volume bails out price.

FICO short — the cleanest disclosed position, and it is not an AI call. The regulatory break (VantageScore acceptance, Pulte, Rocket) is public and has already moved the stock. Whether share goes from a small base toward the 50% he floated is the remaining test.

13F cross-check — unavailable. There is no 2026 institutional filing to confirm size, timing, or the four hedged names. His book is whatever he says on air. The Google sale and the FICO short are the only positions with a clear public timestamp and a stated reason. The AI hedge is real as a risk-management description and unverifiable as a holding list.

For someone trying to use this framework: the tradeable distinction he is drawing is between a commoditizing model layer and a still-bottlenecked infrastructure layer whose demand is downstream of two loss-making customers with take-or-pay contracts. The data through early October 2026 confirms the commoditization and the contractual concentration. It does not yet confirm that open-weight share, or a lab funding break, has been large enough to reverse infrastructure orders. That is the claim still open.

---

Recent Findings Supplement (October 2026)

Steve Eisman has consistently argued since mid-2026 that frontier AI model providers (primarily OpenAI and Anthropic) lack durable economic or pricing moats, as users switch freely between models and cheaper open-weight/Chinese alternatives erode pricing power, setting the stage for price wars that could cascade through the ecosystem.[1][2]

This view has sharpened in recent months (September–October 2026), with Eisman describing AI labs as “manufacturing a crisis” around safety/doomsday scenarios to pursue regulatory capture and duopoly protection, while noting that “token maxing is over” and open-weight models are gaining share. He sees the entire AI supply chain—from chips to hyperscalers—as structurally dependent on just two private companies remaining healthy.[3][4]

In October 2026 interviews, he revealed he has begun partially hedging four AI-related holdings by shorting portions “against the box” over the past month (a shift from his August stance of not shorting the trade), while remaining “mostly long” overall and calling a major implosion call “premature.” His core concern is concentration risk rather than valuations.[1][1]

Named Positions and Rationale (2026 Developments)

Eisman has named or implied the following in 2026 commentary (primarily June–October):

• Nvidia (NVDA) — Infrastructure long/pick-and-shovel favorite: Preferred over hyperscalers because it supplies essential chips that must be bought regardless of which model wins; cited strong revenue growth (e.g., ~85% YoY in one quarter referenced) and ecosystem role, though he flagged customer concentration risk (top five customers ~70% of AR in recent filings).[5][5]
• Cisco (CSCO) and Arista (ANET) — Positive infrastructure exposure: Highlighted Cisco’s AI-driven acceleration (18% revenue growth reported, $4B hyperscaler AI orders in a recent quarter, hyperscaler business on track to nearly double in FY2027) as evidence the boom is “for real,” while positioning networking gear as resilient picks-and-shovels plays alongside Nvidia.[6][6]
• OpenAI and Anthropic (private) — Central concern/Achilles’ heel: OpenAI described as the “weaker” of the two; together they purportedly drive ~70% of AI-related revenue at the hyperscalers (25–35% of their total cloud revenue). Any major issue (e.g., financing crunch, bankruptcy) could unravel the chain. Cited Anthropic’s disclosed 2025 revenue (~$4.6B) vs. >$8B operating loss and $518B future infrastructure obligations.[1][7]
• Hyperscalers (MSFT, AMZN, GOOGL, ORCL) — High exposure/structural vulnerability: Named as having heavy reliance on the two labs; compared to capital-intensive airlines with little pricing power due to easy model switching. Specific flags include Oracle’s reported backlog concentration and Alphabet’s massive equity raises to fund AI capex.[8][5]
• Other notes: Would not invest in a potential Anthropic IPO; earlier comments panned SpaceX valuation as excessive (non-core to AI thesis). Partial hedges applied to four unnamed AI holdings in recent weeks.[9]

No comprehensive personal 13F filings are publicly tied to Eisman (he is described as ex-Neuberger Berman in recent coverage); Neuberger’s Q2 2026 13F shows large positions in NVDA, AMZN, MSFT, and GOOGL, consistent with broad AI exposure but not directly attributable.[10]

API Pricing Trends and Inference Cost Declines (Supporting Data)

Recent months show concrete evidence aligning with Eisman’s price-war thesis:
- OpenAI cut Luna pricing ~80% in July 2026 (driving ~10x usage increase per its CFO); further cuts on GPT-6 Sol/Luna models in September.[6][11]
- Multiple labs (OpenAI, Anthropic, others including Chinese/open-weight providers) implemented aggressive reductions in September 2026, with frontier and mid-tier prices falling sharply (e.g., blended drops of 33–64% in some lines since July).[11][12]
- Broader inference costs: Average ~$1.16–1.18 per million tokens by early August 2026 (43% decline in ~10 weeks from late May); Epoch AI tracks fixed-capability costs falling ~47% per quarter (~13x annually) on average, with ranges of 9–900x depending on task/benchmark. Some equivalent GPT-3.5-level performance dropped from ~$20/M tokens (2022) to fractions of a cent.[13][14]

These trends validate the mechanism Eisman described: competition from open-weight/Chinese models and efficiency gains forcing price competition, squeezing model-provider margins while potentially boosting volume.

Testable Claims and Data Alignment

• Price wars and eroding moats: Strongly supported by documented multi-lab price cuts, usage elasticity to lower prices, and open-weight share gains. Claims are directly testable via public API pricing pages and usage reports.[13]
• Concentration risk (~70% figure): Directional risk is plausible given disclosed Anthropic metrics and hyperscaler capex/backlog commentary, but the precise percentage lacks an independent public source breakdown and remains an Eisman estimate.[1]
• Hedging actions: Observable in principle via portfolio disclosures or statements; his October comments confirm recent risk reduction on AI names.
• Overall durability: The boom’s “for real” demand side (e.g., Cisco orders) is backed by company reports, while the moat/price-war downside is evidenced by pricing trends. Claims about regulatory motives or exact failure probabilities are more interpretive.

For competitors or entrants: Eisman’s framework prioritizes verifiable infrastructure suppliers (chips/networking) with sticky demand over application-layer model or hyperscale plays exposed to commoditization. Monitoring token pricing indices, hyperscaler customer concentration disclosures, and open-weight benchmark performance provides ongoing tests of the thesis. Recent hedging signals his own view that risks have risen without invalidating the secular demand story.

Eisman’s recent comments are concentrated on lab concentration and manufactured-crisis regulation. I’m now pulling disclosed positions and the depreciation, circular-deal, and power arguments from the other skeptics.
Eisman is a concentration skeptic who is still mostly long. Burry is short the accounting. Chanos is short the landlords. Zitron says the cash never arrives. Marcus says the technology commoditizes before the capex pays back. They share a fragile center—OpenAI and Anthropic—and split on whether Nvidia is the safe side of the trade.

Eisman’s actual argument is not “AI is fake”

Steve Eisman’s 2025–2026 case is a credit-and-customer argument, not a technology-doesn’t-work argument and not a valuation-multiple argument. On October 2, 2026 he said his main worry is concentration, not stock prices: roughly 70% of AI-related revenue at Microsoft, Amazon, Alphabet, and Oracle comes from OpenAI and Anthropic, and those two labs are 25% to 35% of the hyperscalers’ cloud revenue. “If anything bad happens to one of those two companies within the next year, everybody’s in trouble.” He has started hedging, but he still calls a full implosion call premature. [1] [2]

The mechanism he keeps returning to is the absence of a pricing moat. Open-weight and Chinese models are taking share, “tokenmaxxing” is over, and a price war at the labs would hit the cloud backlog that the public companies have already capitalized. He treats the labs’ extinction warnings as a regulatory strategy: manufacture a crisis, get rules they can shape, and rebuild the duopoly that open models are eroding. “Terminator” risk, in his words, is “garbage,” and there is “no evidence” current systems are near AGI. [3] [4] [5]

His sharper 2026 financial point is structure, not useful-life math. On his podcast he said off-balance-sheet techniques are “back with a vengeance.” Meta’s Louisiana Hyperion campus is financed through Beignet Investor LLC: Meta owns 20%, Blue Owl-managed funds the rest, the vehicle issued about $27.3 billion of debt, and Meta can lease the site for up to 20 years while bearing delay and overrun costs—with the project debt kept off Meta’s balance sheet. Ernst & Young flagged the variable-interest-entity judgment as a critical audit matter. Oracle, in his telling, did the opposite and paid for it: $43 billion of debt raised in fiscal 2026, free cash flow of negative $23.7 billion, an S&P cut to BBB−, and Project Jupiter debt trading around 90 cents on the dollar. [6] [6] [7]

Disclosed book. Mostly long. Over the month before October 2 he “shorted part” of four unnamed AI holdings against the box—partial hedges, not a directional short book—and said he was not recommending the tactic. In August he was explicitly not shorting the trade. In June 2026 he preferred the infrastructure layer (Nvidia, Arista, Cisco) over hyperscalers precisely because service providers have no moat. In July he exited Alphabet after capex guidance as high as $205 billion and a negative free-cash-flow quarter. He has said he would not buy an Anthropic IPO. [1] [8] [9]

What this means if you are trying to trade against him: he is not Burry. A depreciation restatement or a Nvidia multiple compression does not, on his own words, break his thesis. A funding or share-loss event at one of the two labs does.

Depreciation: Burry’s tell, Eisman’s shrug, Chanos’s ROIC knife

Michael Burry’s load-bearing claim is that hyperscalers are stretching GPU lives to five or six years when the economic life is closer to two or three, suppressing depreciation by about $176 billion from 2026 through 2028. He has called extending useful life “one of the more common frauds of the modern era.” Secondary write-ups put the 2028 earnings overstatement near 27% at one of Oracle or Meta and near 21% at the other; those write-ups disagree on which name gets which figure, so the company-level split should not be treated as settled. [10] [11] [12]

Eisman looked at that math in November 2025 and rejected the conclusion. He granted the arithmetic and said he did not think the concerns “matter that much.” The question that matters, in his framing, is whether the spending produces returns and cost savings—and “we won’t know the answer for a while.” By August 2026 he was watching Meta’s depreciation as a real P&L weight (server and network depreciation of $4.62 billion in the quarter, up about 48%), not as a fraud flag. That is the cleanest documented disagreement between the two Big Short investors. [13] [14]

Jim Chanos uses depreciation as a returns test, not an earnings-quality indictment. In October 2025 he took CoreWeave’s own “6 or 7 years” rental claim, amortized about $19.1 billion of capital employed over seven years, and got roughly $2.73 billion of economic depreciation against about $2.62 billion of annualized EBITDA—a 0% return on invested capital at the generous end of management’s life assumption. His broader line is the telecom accounting mismatch: Nvidia books the chip sale as profit now; the buyer capitalizes it and expenses it over 5 to 10 years, so a capex boom inflates S&P operating earnings until orders stop. He has also said he is not the two-year-life straw man—he has run the neocloud short on a 10-year life and still gets mid-single-digit pretax returns. [15] [16] [17]

Harris Kupperman (Praetorian Capital) sits with Burry on lives and with Zitron on the revenue gap. He first assumed a 10-year blend on a roughly $400 billion 2025 data-center spend and concluded the industry needed on the order of $480 billion of revenue for an adequate return, against something like $15–20 billion of AI revenue. After talking to operators he shortened the life—buildings and power systems as well as chips, three to ten years—and said he had not been bearish enough. [18] [19]

Where they land versus Eisman. Burry and Kupperman think the books are already lying about cost. Chanos thinks even honest long lives do not produce a good business in the middle of the stack. Eisman thinks the schedule is a second-order issue next to whether OpenAI and Anthropic can pay for what they have promised to consume.

Circular deals: same diagram, different crime

The shared picture is a loop: a chipmaker or cloud funds a lab, the lab commits to buy compute, the cloud books revenue and orders more chips, and equity marks rise on the commitments. The skeptics disagree about what the loop proves.

• David Einhorn gave the cleanest unit example in Greenlight’s Q3 2025 letter: $1 of loss-making ChatGPT spend can cascade into more than $8 of reported AI revenue across OpenAI, Microsoft, CoreWeave, and Nvidia, then into $100–$200 of market wealth. He cited McKinsey’s $6.7 trillion global data-center spend through 2030 and said he was refusing to participate. His point is capital destruction even if the technology works. [20] [21]
• Burry has called the web “a picture of fraud, not a flywheel,” with “true end demand ridiculously small” and “almost all customers funded by their dealers.” He has also tallied nearly $1.2 trillion of uncommenced lease commitments and more than $1.5 trillion of purchase commitments at the five hyperscalers—obligations that do not show up as ordinary debt. [22] [10]
• Zitron uses the same CoreWeave loop Eisman worries about in SPV form: Nvidia funded CoreWeave, became a major customer, and CoreWeave borrowed against the contract and the GPUs to buy more GPUs. His September 2026 break-even claim is that hyperscalers need about $308 billion a year of AI revenue to cover 2026–27 capex and have about $183 billion, 64% of it from Anthropic and OpenAI. [23] [24]
• Eisman says the circular-financing debate is real and “disturbing,” but he ranks the off-balance-sheet vehicle as the smell test that fails. Concentration does the rest: if the two labs are the end demand, the loop is not a diversified flywheel. [6] [25]

Reuters’ look at Anthropic’s confidential prospectus is the fact that makes Eisman’s concentration claim falsifiable. Revenue rose about twelvefold to nearly $4.6 billion in 2025; the operating loss was about $8.06 billion; compute and infrastructure cost $7.33 billion. The headline net loss near $42 billion includes roughly $34 billion of non-cash remeasurement on financing that can convert into shares. Future cloud and infrastructure obligations were put at $518 billion. The company is seeking a valuation around $2 trillion. Forty-seven percent of 2025 sales were routed through Amazon and Google, which are also investors, suppliers, and rivals. [26] [27] [28]

ROI and valuation: “we won’t know” versus “the math already failed”

Eisman’s ROI stance is deliberately unfinished. In November 2025 the returns question was the one that mattered and the one that could not yet be answered. In August 2026 he called Meta’s cost curve “astonishing”—expenses up 55% against 28% revenue growth, R&D up 67%—and treated exploding depreciation as a weight, not a verdict. In October 2026 he still would not make the implosion call. [13] [14] [1]

The others treat the gap as already decisive.

Ed Zitron’s line is that ROI cannot be measured because it is not there. Enterprises moved off all-you-can-eat plans onto token billing and, in his reporting, could not show a return once they paid something closer to cost. He has described generative AI as a roughly $50 billion revenue industry “masquerading as a one trillion-dollar one,” and Investor’s Business Daily put the market-value added in AI-linked stocks since late 2022 at $27 trillion. His October 2, 2026 warning shifted from narrative to credit: private credit funding the buildout is a “brewing crisis,” Goldman Sachs counts $88 billion of lower-rated AI-related borrowing this year, and CoreWeave reported $35.6 billion of debt as of June 30. [29] [23] [30] [31]

Gary Marcus attacks the same gap from the product side. LLMs, in his view, have no technical moat, hallucinate, and are being priced as commodities, so token price wars destroy the margins the capex model needs. On September 30, 2026 he said Anthropic’s ARR looked like it was flatlining “post tokenmaxxing” just as the company headed toward an IPO, and that a $2 trillion market cap does not survive an expected-value test once competition, open source, local models, and liability are admitted. He has also said Nvidia eventually declines once that commoditization is recognized—directly opposed to Eisman’s 2026 preference for the chip and networking layer. [32] [33] [34] [35]

Chanos’s valuation point is relative, not absolute. Neoclouds and ex-bitcoin miners are equipment-leasing companies. They should not trade at higher multiples than Nvidia, AMD, or TSMC, which control supply. Established data-center operators, in his telling, earn mid-to-low single-digit pretax returns—“a really bad business.” He has been short that cohort since 2022 and said in late 2025 he was doubling down. [36] [37]

Paul Kedrosky, who shares Chanos’s “four forces” frame (technology, real estate, credit, policy), adds the price path: tokens are a hyper-deflationary commodity, with performance-adjusted prices falling on the order of 70–80% a year, while the debt used to build the factories is long. More than 60% of AI financing is now debt-backed, up from roughly 15–20% a year earlier, on his September 2026 estimate. Usage has to rise hundreds of percent forever just to stand still. [38] [39] [40]

Power is a timing constraint, not a separate religion

Eisman’s power point is mostly about time and credit ratings: data centers are slow and expensive, Nvidia has raised chip prices, and companies hide the debt so the rating does not move. He has also flagged the 10-year Treasury yield above 5% as a level that could force a correction, because the buildout is now a borrower. [41] [25]

Zitron makes power the operational failure mode. In October 2025 he contrasted OpenAI’s 10-gigawatt Stargate pledge with Abilene’s then-roughly 350 megawatts of generation and a 200-megawatt substation. By mid-2026 his claim had shifted to chips that cannot be turned on: on the order of $200–300 billion of GPUs in warehouses or unpowered halls, and an estimate that roughly half of AI chips sold since 2023 were not installed. Oracle’s force majeure notice on Project Jupiter—the same project whose debt Eisman watched trade at 90 cents—is the credit version of that delay. [23] [42] [43]

Kedrosky and Chanos treat the grid as the reason the real-estate leg of the bubble is unstable. Data centers have asked Texas’s grid for 435 gigawatts against a system that has never delivered more than 85 gigawatts at once, while plants take six to eight years and halls take about two. Chanos’s jab at “data centers in space” is that if orbital compute is the answer, the terrestrial build coming online in two to three years is a terminal short—and if it is not the answer, the narrative itself is late-cycle decoration. [44] [45]

Comparison matrix

Issue	Eisman	Burry	Chanos	Zitron	Marcus	Einhorn / Kupperman / Kedrosky
Depreciation	Not the tell. Nov 2025: Burry’s schedule math “doesn’t matter that much.” Later treats Meta’s rising depreciation as a real cost, not fraud. [13]	Core claim. 2–3 year economic life vs 5–6 year books; ~$176B understated depreciation, 2026–28. Calls life extensions a common modern fraud. [10] [12]	ROIC knife. CoreWeave ~0% at a 7-year life. Capex boom inflates earnings because sellers recognize profit and buyers defer it. Will also run the short on a 10-year life. [15] [16]	Secondary to cash. Uninstalled or unpowered GPUs mean depreciation starts before revenue does. [42]	Not his frame. Obsolescence shows up as price wars, not GAAP lives. [32]	Kupperman: 10-year blend was too kind; 3–10 years, and 2025 capex already needs hundreds of billions of revenue to break even. [18]
Circular deals	Real and disturbing, but ranked behind SPVs and two-customer concentration. [6]	“Fraud, not a flywheel.” End demand “ridiculously small.” Also ~$1.2T leases + $1.5T purchase commitments. [22] [10]	Vendor-financing rhyme with Lucent/Nortel and 1999–2000. Customers doing the spending are unprofitable—worse than telecom. [46] [47]	The mechanism. Nvidia–CoreWeave–debt loop; OpenAI–Oracle commitments. ~64% of AI revenue from the two labs. [23] [24]	Implied: labs must keep raising because the product does not fund the commitments. [48]	Einhorn: $1 of user spend becomes $8 of reported AI revenue. Kedrosky: the loop makes true demand unreadable. [20] [49]
ROI	The open question. Meta’s cost curve is already wrong-way. A definitive failure call is still early. [14]	Write-offs follow when utilization is confused with economic benefit. [22]	Neoclouds are low-single-digit ROIC equipment lessors. “Long what the chips produce, not where they reside.” [17]	No measurable ROI. Token billing exposed subsidized pricing. Cash flow, not revenue, is the constraint. [29] [31]	Tokens burned “without any real significant ROI.” Margins go to commodity levels. [32]	Einhorn: even a world-changing technology can destroy the capital spent this way. Kupperman: depreciation already exceeds revenue. [50] [19]
Valuation	Not the lead risk. Will not buy Anthropic. Has preferred Nvidia/networking over hyperscalers. [8]	Puts struck far below market (Nvidia Sept 2027 puts in the mid-$100s vs about $229). [51]	Middlemen should not out-multiple TSMC, Nvidia, or AMD. SpaceX-style “hopes and dreams” IPOs are a late-cycle tell. [36]	$27 trillion added since late 2022 on a business that does not cover its build. [30]	Anthropic at ~$2 trillion and a rising OpenAI mark fail expected-value math. Nvidia eventually rerates down. [33] [35]	Einhorn sitting out. Kedrosky: required returns of 10–12% on stressed projects cannot be met if token prices keep falling. [52] [39]
Power / time	Slow builds, higher chip prices, incentive to hide debt. 10-year yield above 5% is a correction trigger. [41]	Less central than accounting and circular demand. 1960s leasing parallel is about residual values, not the grid. [53]	Data-center real estate is one of four simultaneous bubble drivers. “Data centers in space” implies the terrestrial book is terminal. [40] [45]	Binding constraint. Stargate power gap; hundreds of billions of chips not installed; delays mean interest compounds before revenue. [23] [43]	Not a grid analyst. Local models are the economic threat, not megawatts. [34]	Kedrosky: Texas requests of 435 GW vs a grid that has never delivered more than 85 GW. Debt wall around 2029; he has also said a break could be 6–12 months out. [44] [39]
Safety / AGI talk	Manufactured crisis to win regulation and a duopoly. No moat, no near-term AGI. [4]	Not his argument.	Debated it with Marcus; his break point is financing and IPOs, not extinction odds. [54]	Models do not do what is advertised; conflating LLMs with all of AI is the sales trick. [29]	Agrees AGI is not close. Disagrees that the only problem is cynicism: liability, agent failures, and bad expected-utility math are real. [55] [34]	Kedrosky: the technology is real and consequential; the financial structure around it is the unstable part. [38]
Disclosed position	Mostly long. Partial against-the-box hedges on four unnamed AI names. Sold Alphabet. Would not buy Anthropic. Some shorts, not a dedicated AI short fund. [1]	Replaced outright shorts with 2026–27 puts on Nvidia, Palantir, Micron, Nebius, Oracle, Caterpillar, SOXX, and Nasdaq 100. Closed CoreWeave short pending better-priced puts. Timeline pulled in from a 2028 base case; “more confident than ever” of a reckoning over the next year. [51] [56]	Short data-center landlords, neoclouds, and miner-to-landlord conversions (CoreWeave, IREN, Cipher named in interviews). Conceptually long the chip producers. Not willing to call a bubble except in hindsight. [37] [57]	No book. Public prediction that the money runs out and the break cascades from a lab or a neocloud, not from a single Nvidia print. [23] [31]	No book. Public critic; told readers not to underwrite a $2 trillion Anthropic. [33]	Einhorn: refusing to participate, not a published single-name AI short in these sources. Kupperman: published the revenue-gap math, not a named short book here. Kedrosky: venture/economist critic, not a short seller. [52]
Timing	Premature to call the bust. Lab stress inside a year, or a 10-year yield stuck above 5%, is enough for a correction. [1]	Sooner than later. Confidence interval is the coming year, not 2028. [51]	2027–28, when unprofitable customers cannot keep funding the spend. IPO wave is the historical warning. [47]	Has said roughly 18 months from October 2025, and more recently that credit stress is already here. [23] [31]	2026 is the year retail and index holders are left with the bag; collapse date unknown. [32]	Kedrosky: 6–12 months on the outside in one September 2026 interview, and a 2029 maturity wall in another. Those are his ranges, not a consensus. [39] [49]

Where they actually agree and disagree with Eisman

Agreement cluster. The two labs are the fragile center of reported AI demand. Circular commitments overstate end-customer demand. Open models and token price wars threaten the revenue that is supposed to service the build. Off-balance-sheet and private-credit structures are how the industry is stretching a cash-flow business into a leverage business. Anthropic’s own prospectus—$4.6 billion of 2025 revenue, an $8 billion operating loss, $518 billion of future obligations, a hoped-for $2 trillion valuation—is the document all of them can point at. [27] [2]

Disagreement that matters for a portfolio.

1. Nvidia. Eisman has treated the infrastructure layer as the relative winner because someone has to buy the chips whoever wins the model war. Marcus says that advantage dies once models are commodities. Burry is outright bearish via puts struck in the mid-$100s. Chanos is the hybrid: long the producers, short the landlords who rent their output. [8] [35] [51] [17]

2. Depreciation as the trigger. Eisman explicitly broke with Burry. If you need an accounting restatement to be right, you are in Burry’s trade, not Eisman’s. If you need neocloud ROICs to be recognized as leasing returns, you are in Chanos’s.

3. Whether the bust call is mature. Zitron, Kupperman, and (as of late September 2026) Burry say the revenue math has already failed and the timing has moved forward. Eisman still says the returns question is unanswered and a major implosion call is premature. That is a position difference, not a wording difference: he is hedging longs; they are trying to be short into the break.

4. What the doomer talk means. Eisman and Marcus agree current systems are not near AGI. Eisman thinks the slowdown rhetoric is a moat strategy and dares the labs to postpone the IPO. Marcus thinks liability, agent failures, and bad expected-value math are real constraints the valuation ignores. Chanos’s break point, on the September 25, 2026 RiskReversal conversation with Marcus, is the financing and IPO window, not extinction probabilities. [4] [54] [34]

The practical split: Eisman fails if the two labs keep raising and cloud revenue stays diversified enough that a price war does not hit reported hyperscaler numbers. Burry fails if six-year GPU lives keep earning their book and his 2027 puts expire. Chanos fails if neocloud lease rates and residual values stay high enough to justify REIT-like multiples. Zitron fails if external, non-lab cash demand shows up before the credit rolled to build the halls comes due. Marcus fails if a durable technical moat appears and token prices stop collapsing. Those are different bets wearing the same “AI skeptic” label.

---

Recent Findings Supplement (October 2026)

Steve Eisman has recently highlighted extreme concentration risk in the AI ecosystem, where roughly 70% of hyperscaler AI revenue (Microsoft, Amazon, Alphabet, Oracle) flows from OpenAI and Anthropic, making the entire trade vulnerable to issues at either lab—particularly the weaker OpenAI.[1][2]

In September–October 2026 interviews and podcasts (e.g., CNBC, Prof G Markets), he described trimming some AI exposure and hedging select positions while remaining mostly long, citing nervousness about the narrative rather than valuations per se. He views OpenAI/Anthropic “doomsday” or “Terminator” safety rhetoric as a manufactured crisis to engineer regulation and create moats/duopolies amid eroding advantages from open-weight models and price competition.[3][4]

Eisman also flagged off-balance-sheet financing via SPVs and guarantees as reminiscent of Enron/GFC-era tactics, warning it obscures leverage in the capex boom.[5][6]

Michael Burry has intensified warnings on hidden liabilities and fundamental limits, estimating ~$3 trillion in uncommenced leases, purchase commitments, and contingent exposures across hyperscalers, often kept off-balance or structured to suppress reported depreciation (potentially by ~$176 billion cumulatively 2026–2028).[7][8]

In late September 2026 X/Substack posts, he moved up his timeline for an AI bubble burst (now expecting it within the next year with high confidence) and shorted or adjusted positions in Nvidia, semiconductors, and related names.[8]

Burry argues LLMs cannot reach AGI/understanding because “understanding cannot exist unless reason first exists without language,” with synthetic data training risking model collapse via error propagation and compression of limited human knowledge.[9][10]

He has called safety slowdown calls by executives self-serving (marketing ahead of IPOs/valuations) and drawn parallels to 1960s computer cycles and 1990s telecom overinvestment.[11][12]

Jim Chanos has framed the AI boom as a unique “four-in-one” bubble (policy support, tech enthusiasm, credit financing, and data-center real estate construction) and characterized data centers/neo-cloud operators as capital-intensive equipment-leasing businesses with declining incremental returns on invested capital (ROIC peaked ~2024, potentially below cost of capital by mid-2027).[13][14]

In a September 2026 podcast with Gary Marcus, he emphasized circular financing risks (e.g., Nvidia guarantees for its own chip sales to projects like OpenAI data centers) and compared the speed/scale unfavorably to the dot-com era, noting sudden pullbacks remain possible even if secular demand persists.[15][16]

Chanos has shorted data centers, neo-clouds, and related plays while hedging broader market exposure.[17]

Gary Marcus has focused on technical and operational shortcomings, arguing LLMs lack durable moats, remain hallucination-prone and commoditizing (driving token price wars that threaten OpenAI/Anthropic profitability), and that “agentic” systems pose acute risks due to poor judgment, inability to reliably follow rules, and excessive system/internet access.[15]

In September–October 2026 Substack posts and commentary, he highlighted rising AI agent security incidents (now tens of thousands, per Axios reporting, with many unknown or potentially harmful) and criticized self-regulation as insufficient “safety theater,” calling for stronger oversight while downplaying extinction risks.[18][19]

He advocates neurosymbolic/world-model approaches over pure scaling and has noted OpenAI’s agent experiments as foreseeable problems ignored internally.[20]

Ed Zitron has amplified debt and sustainability concerns, warning that private credit and complex structures (e.g., Amazon/CoreWeave vehicles moving ~$8B in Nvidia chips into leased-back arrangements) represent a “brewing crisis” amid delays, higher borrowing costs, and >$1T already spent on buildout with another trillion eyed.[21]

In recent interviews (Bloomberg, etc.) and his newsletter, he argues capex primarily props up two unprofitable labs (OpenAI/Anthropic) rather than broad demand, labels much of the narrative a “lie,” and views safety efforts as theater.[22]

Zitron’s earlier timeline predictions (bubble by Q2 2026) have not materialized, but he notes growing media acknowledgment of his economic critiques.[23]

Agreements across the group (Eisman, Burry, Chanos, Marcus, Zitron) center on unsustainable economics: heavy reliance on OpenAI/Anthropic, circular/off-balance-sheet financing and debt opacity, eroding or absent moats/ROI, and self-serving elements in safety/regulation talk. They converge on valuation and power/concentration risks, with recent disclosures underscoring Anthropic’s ~$4.6B 2025 revenue against >$8B operating losses.[1]

Disagreements are mainly in emphasis and timing: Eisman sees real underlying demand but acute concentration vulnerability and views full implosion as premature; Burry and Chanos are more aggressively bearish on imminent bubble dynamics and hidden leverage/depreciation; Marcus prioritizes inherent technical unreliability and agent dangers over pure economics; Zitron stresses outright deception in the capex narrative. Investor positioning varies—Eisman mostly long with hedges; Burry and Chanos actively short AI-related names; Marcus and Zitron are primarily commentators.[1][7]

Comparison Matrix (Core Arguments, Recent Positions, Alignment with Eisman – Post-April 2026 Developments Only)

• Depreciation Accounting / Hidden Liabilities: Burry (suppress ~$176B via structures, $3T total commitments); Eisman (SPV/off-balance akin to GFC); Chanos/Zitron (complex private credit/leasebacks). Strong agreement with Eisman on opacity.[6][8]
• Circular Deals / Financing: All highlight (Eisman SPVs/guarantees; Burry/Chanos Nvidia backstops for own sales; Zitron private credit vehicles). Near-universal agreement.[5][16]
• ROI Gaps / Returns: Chanos (hyperscaler incremental ROIC declining toward/below WACC by mid-2027); Zitron (unsustainable for labs); Burry (spending vs. economics mismatch). Aligns with Eisman’s durability doubts.[14]
• Valuation / Bubble: Burry/Chanos/Zitron (imminent burst risk, four-in-one or rot-com elements); Eisman (nervous but mostly long, concentration > valuation). Partial overlap.[8][13]
• Power / Concentration Constraints: Eisman (70% hyperscaler AI rev from 2 labs); others echo dependency and capex scale. Strong shared concern.[1]

Implications for competitors or entrants: These converging critiques—especially financing opacity, concentration, and ROI compression—suggest any new AI infrastructure or model play must demonstrate transparent economics, diversified revenue beyond the two labs, and defensible technical differentiation (e.g., neurosymbolic per Marcus) to attract capital amid rising skepticism.[21]

Sources are limited to post-April 2026 publications (primarily Sept–Oct 2026). No comprehensive head-to-head debates surfaced beyond the Chanos/Marcus podcast; claims rely on direct statements from the individuals.

Eisman’s 2026 case is not “AI is fake.” It is that the stack is a two-name credit chain with no pricing moat at the model layer. The strongest bulls answer pieces of that—and systematically talk past the piece that would actually break the chain.

What Eisman is actually claiming

On Prof G Markets (October 2) and earlier CNBC and podcast appearances, Steve Eisman said his worry is concentration, not valuation. He estimates roughly 70% of AI-related revenue at Microsoft, Amazon, Alphabet, and Oracle comes from OpenAI and Anthropic, equal to about 25–35% of those firms’ cloud revenue. “If anything bad happens to one of those two companies within the next year, everybody’s in trouble.” He calls OpenAI the weaker of the two, and he has started hedging part of four AI holdings “against the box” rather than shorting the trade outright. He has not sourced the 70% figure. [1] [2] [3]

The rest of the case has three legs:

• Receivables and backlog concentration. Note 7 of Nvidia’s 10-Q, as he reads it, shows the top five direct customers at about 70% of accounts receivable. He also says roughly half of Oracle’s reported backlog is OpenAI. [4] [5]
• No model moat, so a price war. Users flip between ChatGPT, Gemini, and Claude. “Tokenmaxxing is over.” Open-weight models are taking share. “There are no pricing moats in this business. Today I have the best LLM and tomorrow yours is better and cheaper.” He reads safety slowdown calls from Dario Amodei and Sam Altman as an attempt to manufacture regulation that would lock in a duopoly the labs cannot earn in the market. [6] [7] [3]
• Obligations vs. cash. A prospectus Reuters reviewed showed Anthropic at nearly $4.6 billion of 2025 revenue, an operating loss of more than $8 billion, and about $518 billion of future infrastructure obligations. He has also argued OpenAI’s costs have been rising faster than revenue, and that off-balance-sheet vehicles (not the disclosed circular Nvidia financing) are the part designed to obscure. [1] [8]

His own qualifier matters: the chain holds as long as the two labs keep growing fast and keep raising money. The break is a funding or share-shift event at one lab, not a slowdown in token curiosity.

Huang: the chip layer is fungible; the credit layer is not his problem

Jensen Huang’s strongest 2026 answers are aimed at the transmission mechanism—open models and custom chips killing Nvidia—not at whether OpenAI can pay for what it has contracted.

On the August earnings call he said AI had reached an inflection: “Its tokens are productive and profitable. Now, compute is revenue.” Asked whether open models threaten growth because so much demand still comes from closed labs, he said the distinction does not matter to Nvidia. Nearly all open models run on CUDA. Closed and open are both growing, and Nvidia is the only supplier indexed to both. At Goldman Sachs in September he repeated that broad reach—frontier labs, neoclouds, enterprises, OEMs—reduces dependence on any one customer or channel, and that supply, not demand, is the constraint: the company can see more than 100% unconstrained growth but is guiding about 70% for fiscal 2028 because it cannot build more. [9] [10] [11]

CFO Colette Kress put a number on diversification. In the July quarter, data-center revenue was $89.0 billion, up 117% year over year. Hyperscale was about $49 billion; the ACIE bucket (AI clouds, industrial, enterprise, sovereign, regional neoclouds) was about $40.3 billion, up 138% year over year. Management’s line is that non-hyperscale is roughly half the data-center business and growing faster. Huang has also said that if one customer no longer needs capacity, another will take it, because the architecture is general-purpose across training, post-training, and inference. [12] [13] [14]

That is a real rebuttal of “Nvidia dies if Chinese open weights win.” It is a weak rebuttal of Eisman’s credit chain. Two caveats sit inside Nvidia’s own disclosures. A later recast of the hyperscale/ACIE split moved several billion dollars with a single customer-classification decision, so “half” is partly an accounting boundary. And on the same call, management said the AI labs Nvidia expects to support with its own balance sheet should be roughly a quarter of the business next year. Partner financing of more than $500 billion, with Huang capping Nvidia’s share of a given project near 25%, spreads the credit—but it does not remove the labs as the offtake. [15] [16]

On pricing power, Huang’s mechanism is tokens per watt under a power cap, not model differentiation. He has described a system-price ladder from Hopper to Blackwell to Vera Rubin (about $18,000, $25,000, and $40,000 in one Goldman recount) and revenue opportunity per gigawatt stepping up with each generation, with gross margin still around 75% in the July quarter. A 15% server price increase, which Dan Ives treated as bullish, is evidence of scarcity at the rack, not of pricing power at the token. Those can coexist. Eisman’s price war is downstream of the chip. [10] [11] [17]

Nadella: the cleanest rebuttal, and it only covers the flow

Microsoft is the only major bull that has put a number directly against “the cloud is two customers.”

On the July 29, 2026 earnings call, Amy Hood said nearly 90% of full-year Microsoft Cloud revenue came from customers outside frontier-model companies. Commercial remaining performance obligations were $678 billion, up 84%. Excluding OpenAI, RPO grew 25%. All of the sequential RPO increase—about $50 billion—came from customers outside the leading U.S. model makers. Azure grew 43% in the quarter and crossed $100 billion for the year. Microsoft 365 Copilot passed 30 million paid seats. By April, Nadella’s AI business was at a $37 billion annual run rate, up 123%, and the count of customers using a trillion tokens a year had quadrupled. [18] [19] [20]

That rebuts a sloppy version of Eisman (“Microsoft’s cloud is OpenAI”). It does not rebut the version he actually uses (“AI revenue, not total cloud, is 70% two labs”). The stock of commitments is still concentrated: fiscal 2026 commercial arrangements with OpenAI were $24.1 billion, about a quarter of Azure, and OpenAI committed to an additional $250 billion of Azure purchases in the restructuring. Around 45% of the January 2026 RPO balance had been tied to OpenAI. New bookings are diversifying; the backlog that has to be delivered is not. [21] [22]

Nadella’s moat answer is more interesting because it concedes Eisman’s premise. After earnings he wrote that Microsoft is separating harness, context, memory, and action space from any one model family, “because every model is substitutable.” On the call he said the company is designing its own models and chips alongside OpenAI and seeing efficiency gains of up to 40%. In September he said the enterprise agent market will surpass the cloud by orders of magnitude. The moat he is selling is distribution (365, GitHub, Dynamics) plus an agent runtime (Foundry), not GPT weights. That is a direct answer to “LLMs have no moat.” It is not an answer to “if OpenAI’s $250 billion commitment wobbles, Azure’s growth math changes.” [23] [24] [20]

Wood: elasticity is the price-war rebuttal

Cathie Wood’s September investor letter is the cleanest attack on “a price war ends the boom.” She argues inference cost is falling more than 99% a year, and that this is good deflation: in 2025 token demand rose about 25-fold while frontier labs posted revenue growth “the likes of which we have not seen.” Her figures: Anthropic’s annualized run rate went from $9 billion in December to $65 billion by July, and ARK estimates it could top $100 billion by year-end with better profitability than expected—more revenue added in seven months than Salesforce added in 25 years. ARK’s infrastructure spend path is about $500 billion in 2025 to about $1.5 trillion by 2030. In late September ARK bought roughly $80 million of Nvidia and sold AMD. [25] [26]

Independent reporting lines up with the revenue half of that claim and complicates the profit half. A prospectus covered by Reuters and PitchBook showed 2025 revenue near $4.6 billion, first-quarter 2026 revenue of $4.73 billion, preliminary second-quarter revenue above $11.5 billion, and a late-July run rate above $65 billion—against an operating loss above $8 billion in 2025 and $518 billion of future infrastructure obligations. OpenAI’s annualized revenue was approaching $70 billion by late September, with enterprise sales more than doubling since July, per a source cited by Reuters. [27] [28] [1]

Wood’s mechanism is right as far as it goes: unit-price collapse has not stopped dollar revenue from exploding, so “cheaper tokens” is not yet “dead industry.” She does not engage the obligation gap. A lab can be the fastest-growing software company in history and still be unable to fund $518 billion of contracted compute if gross margin does not inflect as fast as the run rate. Celebrating Anthropic’s growth is evidence Eisman himself says keeps the chain intact. It is not evidence the chain can survive a funding stop.

Ives: scarcity and the next TAM, not the two-name risk

Dan Ives has not engaged Eisman’s concentration math. His 2026 bull case is a shortage plus a second demand wave.

On CNBC he put chip demand versus supply at 12-to-1, “physical AI hasn’t even started,” and called the revolution the third inning. By late August he was at “upwards of 15 to one,” arguing a 15% Nvidia price increase would be eaten by enterprises and tokenization, not by a demand air pocket, with equilibrium not until mid-to-late 2028. In late September he told Yahoo Finance physical AI is Nvidia’s “holy grail,” with robotics spending potentially a multiple of today’s infrastructure spend. Nvidia has said physical AI already produced more than $9 billion over the prior twelve months. On October 1 he framed the next three to four years as $4–5 trillion of spending, less than 15% complete, with a $5–6 revenue multiplier per dollar of capex across the rest of tech. An Asia supply check he cited was 13-to-1. [29] [17] [30] [31]

This talks past Eisman. A 12-to-1 shortage explains why Nvidia can raise prices while token prices fall. It does not identify who the marginal buyer of the next gigawatt is. If that buyer is still two labs writing multi-year checks they fund in the private markets, scarcity and concentration are the same fact seen from opposite ends of the rack. Physical AI is a genuine diversifier only if it becomes a large share of orders before a lab-funding break. At a bit over $9 billion against an $89 billion data-center quarter, it is a seed, not a hedge.

Sell-side: they relocated the moat, and some of them restated the risk

The institutional bull case in 2026 does not defend LLM pricing power. It moves the moat.

Morgan Stanley’s “ten laws” frame treats code as commoditized and names three durable advantages: unified data, domain depth (compliance, implementation, workflow embed), and distribution built over decades. Goldman’s September Silicon Valley field work reached the same place: proprietary, continuously updated data that models cannot generate themselves; workflow execution rather than information delivery; pricing shifting from seats to usage, transactions, and outcomes. That is an agreement with Eisman on weights, and a disagreement on whether the industry therefore has no moat. The moat, in this telling, sits in the application and data layer, and in the capital required to build a gigawatt. One buy-side note put the logic bluntly: securing compute costs on the order of $200 billion a year, only four or five firms can self-fund that, and the spending critics flag is the moat. The same note flagged backlog concentration—roughly half tied to OpenAI and Anthropic—as the primary risk, then read it bullishly: the two labs are placing multi-year orders only those vendors can fill. [32] [33] [34]

Goldman’s Ryan Hammond is closer to a qualified concession than a rebuttal. Hyperscalers need about $300 billion of AI revenue in the next few years to break even on the buildout; for solid returns, and for the app layer to earn margin on its compute bill, users need to spend roughly $1 trillion a year on AI applications. Consensus capex for the group is about $806 billion in 2026; Goldman’s analysts are at $1.2 trillion in 2027 and $1.4 trillion in 2028. Cloud revenue is annualizing about $70 billion above the pre-AI trend, and announced backlogs exceed $1.5 trillion. Hammond’s point is that the hurdle is high and years away, not that it has been cleared. A separate Goldman view does not expect supply and demand to balance until the first half of 2028, and notes the median AI-infrastructure stock has already de-rated from 32 times forward earnings in April 2026 to 22 times. [35] [36] [37]

Morgan Stanley’s financing work cuts the other way. One estimate puts 2025–2028 global data-center capex near $2.9 trillion, with hyperscaler cash flow covering about $1.4 trillion and a gap near $1.5 trillion to be filled by private credit, bonds, and securitizations. That is adjacent to Eisman’s off-balance-sheet worry, not a rebuttal of it. The same firm’s return math—GPU leasing around a 31% incremental ROIC, model APIs on owned infrastructure higher, if lease prices hold—says the revenue targets are reachable only if token and rental prices do not crack. Eisman’s whole point is that open weights are the thing that cracks them. [38]

Barclays’ Tom O’Malley raised a Nvidia target to $275 on hyperscaler revenue of $237 billion in 2026 and $401 billion in 2027, with Nvidia’s capture of top-hyperscaler capex climbing toward 44% by 2027. That is a bull case that leans into concentration. It is the mirror image of Eisman, not an answer to him. [39]

Volume evidence the bulls actually cite, beyond lab ARR, is large and mostly not in dispute: Google token processing up roughly sevenfold in a year to multi-quadrillion monthly levels in some 2026 disclosures; Microsoft at over 100 trillion tokens in an earlier fiscal quarter; OpenRouter agentic usage from about 0.5 trillion tokens in February to about 7.3 trillion by August; enterprise generative-AI spend from $11.5 billion in 2024 to $37 billion in 2025 (Menlo Ventures). A quality-adjusted price index found posted prices down only about 9% a year, because buyers upgrade capability rather than buying the same model cheaper. Volume is winning the dollar pool. That rebuts “deflation has already destroyed revenue.” It does not rebut “the dollar pool is still too small, and too concentrated, relative to contracted infrastructure.” [40] [40] [41]

What directly rebuts Eisman, and what talks past him

Eisman point	Direct rebuttal	Talks past it
70% of hyperscaler AI revenue is two labs	Hood: ~90% of Microsoft Cloud revenue, and all sequential RPO growth, is outside frontier labs. Scope mismatch: she measured total cloud and the flow of new bookings, not the AI-revenue slice or the stock of the $250 billion OpenAI commitment.	Ives’ 12-to-1 shortage; Huang’s $3–4 trillion by 2030; Nadella’s “agents dwarf the cloud.” None names the two-lab share.
Nvidia top-five receivables ~70%	Kress/Huang: ACIE ~$40 billion, ~half of data center, growing faster; architecture is fungible across customers; indexed to open and closed models.	Barclays raising targets because hyperscaler capture is rising toward 44%. That confirms the concentration.
Open weights start a price war that reverses the chain	Huang: open models run on Nvidia and expand the installed base; the likely upgrader to Anthropic or OpenAI is someone who already uses AI. Wood: 99% cost drop coincided with 25x tokens and Anthropic $9 billion to $65 billion ARR.	Physical-AI TAM and “third inning.” A second market does not pay this year’s power bill.
Labs have no moat, so capex has no duration	Nadella, Goldman, Morgan Stanley: moat moved to distribution, proprietary data, workflow embed, and the capital required to build a gigawatt. Huang: CUDA plus supply-chain commitments plus tokens per watt.	Wood’s macro “good deflation” and GDP acceleration. True or not, it does not create switching costs at the weight layer.
Obligations dwarf cash ($518 billion; losses; SPVs)	Lab ARR itself—Anthropic above $65 billion by July, OpenAI nearing $70 billion by late September—is the evidence Eisman says keeps the chain from breaking. Early signs of margin inflection at Anthropic are the part of Wood’s note that would matter if confirmed in an S-1.	“Visibility” into land, power, and shells. Booked factories are not collected receivables. Morgan Stanley’s ~$1.5 trillion funding gap restates the worry. Goldman’s $300 billion / $1 trillion ROI hurdle says returns are unproven.

The pattern is consistent. Bulls win the argument about whether demand exists. Token volumes, lab ARR, Azure growth outside the frontier on the margin, and a non-hyperscale half of Nvidia’s data center are all evidence Eisman’s “nothing is happening” straw man never required. They lose, or decline, the argument about correlation. Hood’s 90% and Kress’s ACIE split are the only figures that change the concentration math, and both measure a broader denominator than the one Eisman uses. Everyone else answers a different question: how big the TAM is, how scarce the chips are, or where the moat will sit once models are commodities.

For someone underwriting the trade, the falsifiers are therefore narrow. Watch whether non-frontier RPO keeps accounting for all of the sequential backlog growth after the $250 billion OpenAI commitment starts to be delivered. Watch whether ACIE’s share of Nvidia stays near half once the labs Nvidia is balance-sheet-supporting become a quarter of revenue. And watch the Anthropic S-1—not the July run rate, which both sides already accept, but gross margin versus the $518 billion obligation. If margin inflects, Wood’s elasticity story starts to pay the bills Eisman says cannot be paid. If it does not, the bulls’ demand evidence and Eisman’s credit evidence can both be true at once.

---

Recent Findings Supplement (October 2026)

Eisman's core concerns (primarily Sept-Oct 2026 commentary) center on extreme customer concentration—roughly 70% of AI-related revenue at hyperscalers (Microsoft, Amazon, Google, Oracle) tied to OpenAI and Anthropic, with OpenAI viewed as the weaker link—combined with absent moats, Chinese open-weight models eroding pricing power, and risks of circular financing or pullbacks if labs falter.[1][2][3] Bulls have responded with updated demand metrics, architectural shifts, and diversification evidence, though some arguments emphasize macro tailwinds or next-wave opportunities more than direct rebuttals of the two-lab dependency.

Jensen Huang has countered concentration and moat fears by highlighting Nvidia's broadening customer base and the demand-stimulating effects of open models, while providing supply-chain visibility to support aggressive long-term growth targets. In September 2026 comments and the Q2 earnings context, he noted that non-hyperscaler customers already represent roughly half of Nvidia's data center business and are growing ~100% annually, diluting reliance on the top hyperscalers.[4] He explicitly defended Chinese open-source models (e.g., DeepSeek, Kimi) as drivers of greater overall AI adoption and thus more Nvidia hardware demand, rather than pure commoditizers.[5]

• Nvidia guided to 70% revenue growth for FY2028 (supply-constrained; demand would support higher), with plans to double chip sales volume in calendar 2027 versus 2026; this reflects "high production ramp mode" as AI shifts from research to profitable products in the last six months.[6][4][7]
• New agent-safety software platform and emphasis on physical AI/robotics as the next phase, alongside a record $150B buyback increase signaling cash-flow confidence.[8][9]
• This directly addresses Eisman's concentration point via diversification data but talks past circular-financing critiques by focusing on ecosystem coordination and long-term commitments rather than lab-specific health.

Dan Ives has framed the buildout as early-stage (third inning or ~15% through a $4-5T cycle) with physical AI as the multiplier that dwarfs current infrastructure spend, while treating OpenAI/Anthropic strength as foundational rather than fragile. In late September 2026 remarks, he called physical AI/robotics Nvidia's "holy grail," with potential spending multiples of today's levels, and described the environment as a "1997 moment, not a 2000 bubble."[10][11]

• Demand-to-supply imbalance cited at 12:1 for Nvidia chips; Anthropic and OpenAI positioned as the "heart and lungs" of the trade, with their IPOs expected to bring beneficial transparency.[11][12]
• Views circular elements as building a "new economy" rather than a risk.[11]
• This rebuts moat/pricing concerns indirectly via scale expectations but engages less with the specific 70% hyperscaler revenue concentration statistic.

Cathie Wood/ARK has cited explosive inference cost declines and token demand surges as evidence of sustainable, expanding usage, projecting massive downstream software spend and GDP acceleration while downplaying open-model threats. Recent ARK commentary (Sept-Oct 2026) highlights AI inference costs falling 99.99% annually and token demand rising 25-fold, enabling $3-7T in potential AI-driven software expenditures (19-56% growth).[13][14]

• OpenAI revenue run-rate cited as surging from ~$20B to $70B in a year; Anthropic from ~$9B to $65B+ (potentially $100B by year-end with higher profitability).[15][14]
• Open models seen as non-derailing (ARK signed letter supporting open weights); regulation is the bigger risk. AI could drive 7%+ real global GDP growth (conservative estimate), akin to or exceeding the Industrial Revolution when combined with robotics/energy/storage.[16][17]
• ARK has actively bought more Nvidia shares while trimming AMD.[18]
• This provides strong demand-evidence rebuttals via token/revenue metrics and cost deflation but addresses concentration more through ecosystem growth than direct hyperscaler dependency analysis.

Satya Nadella has pointed to accelerating enterprise adoption metrics and the emergence of agentic workloads as proof of durable, diversified demand, with governance features creating enterprise moats. In September 2026 interviews and earnings context, he highlighted Azure surpassing $100B annual revenue (43% YoY growth in Q4 FY2026), driven by AI/agent consumption on the Foundry platform.[19]

• Customers consuming 1 trillion tokens annually quadrupled YoY; joint Microsoft services customers up 60%; Foundry + Fabric customers up 60%. Agent 365 saw rapid uptake (40M agents in two months).[20]
• AI agents positioned as a market "orders of magnitude" larger than cloud; 2026 flagged as the shift from pilots/experimentation to scaled productivity impact.[19][21]
• Emphasis on trust, auditability, policy controls, and sovereignty options (region-specific models, customization) as enterprise differentiators.[22][23]
• This directly counters moat and pricing-pressure concerns with platform/governance advantages and token/enterprise metrics, while showing Microsoft-level demand resilience that indirectly supports the broader stack.

Sell-side voices like Morgan Stanley have emphasized Nvidia's diversified and higher-growth non-hyperscaler exposure plus improving revenue-per-gigawatt economics from new architectures. A October 2026 reinstatement as top pick noted that while 90% of investor focus is on hyperscalers/frontier labs, half the business (likely the faster-growing half) comes from a broader set of AI model companies, neoclouds, sovereign entities, and enterprises; the 2028 Feynman architecture is expected to lift revenue per gigawatt from ~$40B to well above $50B.[24] Valuation described as undemanding at ~15x FY2028 estimates, with scope for multiple expansion.

Overall mapping: Direct rebuttals appear in Huang's non-hyperscaler stats and open-model defense, Nadella's token/enterprise growth and governance moats, and ARK's cost/token/revenue explosion data—these engage Eisman's concentration and pricing points with counter-evidence. Broader arguments from Ives (physical AI multipliers, early innings) and some Huang/ARK macro/GDP framing talk past the specific two-lab hyperscaler dependency and circular-financing mechanics, focusing instead on forward optionality. Lab revenue updates (OpenAI nearing $70B ARR with surging enterprise; Anthropic quarterly outperformance) provide fresh demand proof points that bulls cite to show sustainability.[25][26]

These developments are concentrated in Sept-Oct 2026 disclosures and commentary; earlier 2026 data showed lower run rates (e.g., Anthropic ~$4.6B for full-year 2025). For competitors or entrants, the bull narrative implies that platforms with strong enterprise controls, agent infrastructure, or exposure to physical AI/robotics and diversified customers are best positioned, while pure hyperscaler or single-lab bets carry the concentration risks Eisman flags.

Eisman’s caution is partly right about concentration and cash burn, and overstated as a near-term systemic call. The labs he treats as fragile are scaling revenue faster than the 2025 loss figures imply, and the hyperscalers are still capacity-constrained with a large non-lab customer base. The price-war risk is real at the model layer, but so far cheaper tokens have expanded usage more than they have killed demand.

What he is actually saying

On CNBC in August and again in early October, Steve Eisman’s AI view is a concentration call, not a valuation call and not a 2008-style short. He estimates OpenAI and Anthropic account for roughly 70% of AI-related revenue at Microsoft, Amazon, Alphabet, and Oracle, and 25–35% of those firms’ cloud revenue. If either lab is impaired within about a year, he says the chain from Nvidia through the hyperscalers reverses. He calls OpenAI the weaker of the two. He has also argued the labs have no durable moat, that open-weight models are taking share, and that safety slowdown talk is a bid for regulation they can shape into a duopoly. He is still mostly long, with partial “against the box” hedges, and has said a call that the AI story implodes is premature. [1] [2] [3]

The 70% figure is his estimate. Secondary write-ups note he has not published a source for the aggregate. Microsoft’s own filing supports a narrower version: OpenAI commercial arrangements were $24.1 billion in fiscal 2026, which some estimates put near 70% of Microsoft’s AI sales, but only about 7% of company revenue and roughly a quarter of Azure. [4] [5] [6]

Lab trajectories undercut “about to break”

The prospectus numbers Eisman cites are real and ugly for 2025. Reuters reporting on Anthropic’s filing showed nearly $4.6 billion of 2025 revenue and an operating loss of more than $8 billion, with about $518 billion of multi-year compute commitments. [1] [7] That is the snapshot his caution is built on. It is already stale as a run-rate picture.

Anthropic’s subsequent reported quarters rewrote the arithmetic. Preliminary Q2 2026 revenue was about $11.5–11.6 billion, roughly 14 times the year-earlier quarter and more than double Q1’s $4.73 billion. Annualized revenue was about $65 billion at the end of July, versus about $9 billion at the end of 2025. The Financial Times reported the company told shareholders adjusted operating income would be positive for a second straight quarter, after a first adjusted operating profit in Q2, with gross margins above 80% before partner revenue share and training costs. That measure strips out stock-based compensation, so it is not free cash flow. It is still the opposite of a company one quarter from being unable to pay its bills. [8] [9] [10]

OpenAI is the cleaner place for Eisman’s “weaker company” label, and even there the revenue path has reaccelerated. Q2 recognized revenue was about $6.7 billion, up only 18% from Q1, while some reports put the operating loss, including stock-based compensation, widening from about $9.3 billion to $12.3 billion. By late September, though, sources told Reuters and Axios that annualized revenue was approaching $70 billion, up more than 70% since the start of the third quarter, with enterprise sales more than doubling since July and Q3 consumer revenue already larger than all of 2025 consumer revenue. Run rate is not audited revenue, and expenses were not disclosed. The direction still contradicts a near-term failure narrative. OpenAI raised $122 billion in March 2026 at an $852 billion post-money valuation, with later reports of talks at $1.2 trillion and above. [11] [12] [13]

What this means for the thesis: a lab can be structurally loss-making on GAAP and still be healthy enough, for the next year, to keep drawing down committed capacity. Eisman’s failure mode requires a break in growth or funding, not merely continued losses. On current reported trajectories, that break is not visible.

Hyperscaler results show demand beyond two customers

Microsoft’s July earnings call is the most direct rebuttal of the cloud-concentration version of the claim. CFO Amy Hood said full-year Microsoft Cloud revenue surpassed $214 billion, with nearly 90% from customers outside frontier-model companies. Commercial remaining performance obligation was $678 billion, up 84%. Excluding OpenAI, that backlog grew 25%, and all sequential RPO growth came from non-frontier customers. Azure grew 43% in the fourth quarter and passed $100 billion for the year. Paid Microsoft 365 Copilot seats were over 30 million. Hood also said cloud demand still exceeded available capacity. [14] [15]

That does not erase OpenAI risk inside the backlog. Earlier in fiscal 2026, OpenAI was reported at about 45% of Microsoft’s revenue backlog after a $250 billion commitment. The point is that recognized cloud revenue and incremental bookings are broader than the two-lab story, and first-party software (Copilot) is a second demand engine that does not disappear if OpenAI stumbles as a renter of GPUs. [16] [4]

Alphabet is the clearest case that AI capex can show up in both growth and margin without being an OpenAI story. Google Cloud revenue rose 82% year over year to $24.8 billion in Q2 2026. Cloud operating income more than tripled to $8.8 billion, and operating margin expanded from 20.7% to 35.6%. Backlog reached $514 billion. Management said nearly 90% of the Fortune 100 uses Gemini Enterprise, and cloud growth accelerated even after excluding the first TPU system sales into customer data centers. Quarterly free cash flow was negative $5.9 billion because capex was $44.9 billion. The cash-flow strain is real. The demand and margin evidence is also real, and it is not a bet that only two startups stay solvent. [17] [18] [19]

Amazon is in the same capacity-constrained regime. AWS revenue was $42.2 billion in Q2, up 37%, the fastest growth in 18 quarters, with backlog of $496 billion. Andy Jassy raised 2026 cash capex guidance to about $220 billion and said even that would not meet 2026 demand, and probably not 2027 demand either. Reported commentary has put AWS AI revenue at a run rate above $25 billion, at least double a year earlier, with custom chips (Graviton, Trainium, Nitro) above a $20 billion run rate. Core AWS usage is rising with AI usage, which is the flywheel a pure lab-customer story misses. [20] [21] [22]

The honest limit: Goldman Sachs has said the five largest U.S. hyperscalers need on the order of $300 billion in annual AI revenue to break even on roughly $800 billion of 2026 AI infrastructure spending, against AI cloud revenue only about $70 billion above the pre-AI trend. Capex is still running ahead of monetization in aggregate. That supports a returns debate. It does not, by itself, support “one lab fails and the U.S. is in recession overnight,” because a large share of recognized cloud revenue is already outside those two names, and several operators say they cannot build fast enough for the demand they already have. [23] [24]

Price cuts have expanded volume; revenue capture is the open question

Eisman’s price-war worry is the strongest part of the bear case, and the data so far cut both ways.

OpenRouter’s window around OpenAI’s late-July cuts to GPT-5.6 Terra and Luna is the cleanest natural experiment. Luna’s daily token volume rose 13.8 times versus its pre-discount average. Terra rose 5.6 times. Sol, which was not cut until weeks later, moved only 1.1 times and functioned as a control. OpenRouter described most of the added volume as net-new usage plus share taken from other providers, not just existing users switching SKUs. Users who adopted during the discount largely stayed, and post-discount daily volume was higher than during the promotion. [25]

The same pattern shows up in aggregate indexes. Apollo’s Torsten Slok noted in June that a token-expenditure index had roughly doubled since late 2025 even as the price of a token had fallen more than 90% since 2023. That is Jevons: unit price down, total consumption up, because new workloads (agents, coding loops, always-on inference) become viable only after the price falls. [26]

The catch, and where Eisman is not obviously wrong, is dollar capture. Goldman’s usage-weighted token price index fell 29% in August alone and more than 50% from its May peak, ending below $1 per million tokens. JPMorgan data cited alongside that report had OpenRouter token usage up about 47% month on month in August while dollar spend rose only about 7%. If price falls faster than volume rises, model-layer revenue compresses even as GPU hours rise. Cheaper intelligence can be good for Nvidia and the clouds (more tokens, more compute) and bad for the labs’ pricing power at the same time. [27] [28]

Enterprise adoption data also splits. McKinsey’s 2026 survey (1,719 respondents) found nearly nine in ten organizations using AI in at least one function, and 44% scaling it across the enterprise, up from 38%. Among firms with at least $1 billion in revenue, 54% said they were scaling, and 40% were scaling agents, up from 27%. The same survey found only 37% reporting a positive EBIT contribution, essentially unchanged from a year earlier. Breadth of use is not the same as proven profit. A price war can coexist with rising seat counts if buyers multi-home and route each workload to the cheapest adequate model. [29] [30]

Track record and the hedged stance

Outside 2008, Eisman’s public record is not a series of correctly timed systemic shorts. The career-defining call was the subprime trade. He has since said, including in a 2024 interview, that he does not expect another financial-system crisis and will not strain to predict the end of the world when he does not see it. That is a useful check on treating every Eisman warning as a Big Short sequel. [31]

The clearest miss is Tesla. He disclosed a short in July 2018 on execution and competition. By February 2020 he had covered, telling Bloomberg that when a stock becomes “unmoored from valuation” and takes on “cult-like” aspects, “you have to just walk away,” and that “there’s no glory in losing money.” Tesla had roughly doubled in 2020 alone by then; short sellers as a group were down billions. He has continued, as recently as 2025 and mid-2026, to call the shareholder base a cult and the operating record a failure. On earnings direction he has often been directionally right. As a trade, the short was wrong, and he said so by covering. [32] [33] [34]

The AI posture has the same optionality, which is why it is easy to criticize as inconsistent. In one breath the labs are nervous, moatless, and manufacturing a crisis because open-weight models are taking share. In the next, if one of them fails tomorrow the U.S. economy is in or near recession and “everybody’s in trouble.” Those claims fight each other. If open models and Google’s own stack are genuine substitutes, a lab failure reallocates spend. It does not zero out inference demand. If they are not substitutes, the “no moat, regulation is the only plan” story is weaker than he presents it.

The position matches the hedge, not the headline. He is mostly long, has trimmed rather than flipped, sleeps fine, and says an implosion call is premature. Against-the-box shorts reduce exposure without forcing a sale. That is a reasonable risk-management choice. It is not a falsifiable market call. If AI keeps working, he was long. If it breaks, he warned. The 2008 trade was the opposite: a large, explicit, money-at-risk short against a consensus that had not priced the risk.

When the thesis fails, and what to watch

The thesis fails if, over the next two to four quarters, three things hold at once.

First, both labs keep converting run rate into cash collections fast enough to service commitments. Anthropic’s adjusted profitability and OpenAI’s reacceleration are the current evidence. A stall in sequential revenue, a down round, or a public cut to contracted compute would revive the call.

Second, hyperscaler AI revenue keeps diversifying. Microsoft’s “90% of cloud revenue outside frontier-model companies,” non-OpenAI RPO growth of 25%, Copilot seat growth, Google Cloud’s 82% growth and 35.6% margin, and AWS saying demand exceeds even $220 billion of capex are the indicators that already lean against a two-customer collapse. Oracle is the exception worth separating: sell-side commentary has put OpenAI at around half of Oracle’s compute backlog, so an OpenAI problem hits Oracle harder than it hits Microsoft’s recognized P&L. [12]

Third, the price war expands total inference spend rather than only reallocating it. Token-volume spikes after price cuts support that. A sustained period in which token prices fall faster than volumes rise, and hyperscaler AI gross margins compress with them, would validate the moat critique even if the labs do not fail.

Indicators that would actually change the odds:

• Anthropic’s S-1: H1 versus H2 growth, gross margin after Amazon and Google revenue share, cash burn versus the $518 billion commitment stack, and customer concentration (the filing already flagged nearly a quarter of 2025 revenue from two clients). [35]
• OpenAI expense disclosure, not another run-rate headline. Revenue near $70 billion annualized with widening operating losses is a funding story, not a self-funding story.
• Microsoft, Amazon, and Alphabet commentary on frontier versus non-frontier mix, and whether backlog growth keeps coming from ordinary enterprises.
• Token price indexes versus token-expenditure indexes. Volume up and dollars up favors the bulls. Volume up and dollars flat favors Eisman’s price war.
• Utilization, not capex. Management already says demand exceeds supply. The bearish turn is idle capacity, guide-downs, or lengthening sales cycles, not another capex raise.
• The 10-year yield, which Eisman himself has flagged: he has said a 10-year that stays above 5% makes a correction probable, because the AI buildout is now partly a credit trade. That is a rates call sitting on top of the concentration call, and it can be right even if the labs keep growing. [36]

Bottom line: the 2025 loss figures and the customer-concentration math are a real risk to the most levered links, especially Oracle and anyone who underwrote multi-year lab commitments as if they were investment-grade demand. They are a weak basis for “everybody’s in trouble” over the next year. The labs’ 2026 revenue paths, Microsoft’s non-frontier cloud mix, Google Cloud’s margin expansion, and the post-discount token surge are the evidence that the cautious view, as a systemic near-term call, is running ahead of the operating data.

---

Recent Findings Supplement (October 2026)

Eisman's core caution centers on extreme concentration risk (roughly 70% of hyperscaler AI revenue tied to OpenAI and Anthropic), potential fragility at OpenAI, and incentives for labs to push safety narratives for regulatory moats amid eroding pricing power. Recent data from mid-2026 onward provides counter-evidence through accelerating revenue at the labs themselves, strong monetization and margins at hyperscalers, rapid enterprise scaling with emerging ROI, and clear volume elasticity from price cuts.[1][2]

OpenAI and Anthropic show explosive revenue growth and funding momentum that undercuts narratives of imminent trouble or over-dependence fragility. OpenAI's annualized revenue run rate approached $70 billion by late September 2026 (up more than 70% since the start of Q3), with enterprise sales more than doubling in that period and Q3 consumer revenue exceeding all of 2025 combined; it is in talks for at least $30 billion more at around a $1.4 trillion pre-money valuation as a bridge to a potential 2027 IPO (following the March 2026 $122 billion round at $852 billion post-money).[3][4]

Anthropic closed a $65 billion Series H in May 2026 at a $965 billion post-money valuation (surpassing OpenAI at the time), with its run rate reaching $47 billion by May and $65 billion by July/August; its 2025 full-year revenue hit nearly $4.6 billion (12x growth from ~$386 million), and a leaked September 2026 IPO prospectus highlighted gross margins swinging to roughly 40% (from deeply negative) alongside $20+ billion in cash despite large operating losses and $518 billion in multi-year infrastructure commitments.[5][6][7]

These trajectories indicate the labs are scaling revenue faster than many expected, with Anthropic appearing ahead on some metrics and both attracting massive follow-on capital despite concentration concerns. For competitors or entrants, this suggests the "two-lab dependency" may prove more resilient than feared if growth continues, but it also highlights the need to monitor actual diversification away from frontier-model reliance and paths to sustained positive free cash flow.

Hyperscaler capex is delivering accelerating cloud revenue and margins, providing validation rather than evidence of unsustainable spend. In Q2 2026, Google Cloud revenue jumped 82% year-over-year to $24.8 billion (accelerating from 63% prior quarter) with operating income more than tripling to $8.8 billion (35.6% margin); Microsoft Azure grew 43% (beating guidance and consensus), with commercial RPO at $678 billion (+84% YoY) and continued H2 acceleration expected; AWS grew 37% (fastest pace in 18 quarters).[8][9]

2026 capex guidance remains elevated (Microsoft ~$175 billion after lease accounting shifts, Amazon ~$220 billion, Alphabet $195–205 billion, Meta $130–145 billion; aggregate hundreds of billions), funded partly by record AI-related debt issuance (~$500 billion by early August), yet cloud backlogs and growth rates suggest demand is absorbing capacity.[10][11] Some analyses flag potential near-term cash-flow pressure as capex trends toward or past operating cash flow in aggregate by late 2026, but profitability and backlog conversion provide a buffer.[12]

This dynamic weakens claims of pure circular or value-destructive spending; instead, it points to AI infrastructure monetizing via higher cloud attach rates and workloads. Entrants should watch Q3/Q4 earnings for margin sustainability and whether growth rates hold as base effects grow.

Enterprise AI adoption data reveals rapid scaling, agentic deployment, and emerging ROI, moving beyond pilots toward production impact. McKinsey's 2026 survey found 44% of organizations scaling AI enterprise-wide (up from 38%), with 40% of large enterprises scaling AI agents (up from 27%); 60% expect higher AI investment.[13] Deloitte's 2026 report noted worker AI access rose 50% in 2025, with the share of companies expecting ≥40% of experiments in production set to double within six months, alongside rising reports of transformative (not just efficiency) impact.[14]

BCG's Applied AI Index 2026 indicated nearly half of companies now generate real value from AI (sharp reversal from prior minimal returns), with "future-built" leaders delivering 2.3x TSR, 2.4x revenue growth, and 2.8x EBITDA growth versus laggards; spending has risen to ~3.3% of revenue.[15] OpenAI's own enterprise signals showed agentic use (e.g., Codex) comprising 64% of tokens among customers by June, with frontier firms generating 8.3x more output tokens per user than typical ones (gap widening threefold since January).[16]

Broader surveys show 73–88% of enterprises using AI in at least one function, though full scaling remains uneven (e.g., only ~7–15% fully scaled in some reports).[17][18] These figures suggest adoption is deeper and more value-accretive than a pure hype narrative allows, particularly in large firms and specific functions like coding/agents. Implications for market participants include focusing on integration/governance tools rather than raw model access, as scaling bottlenecks shift from tech to org design.

Price cuts are driving substantial usage increases that support overall spending and revenue potential, consistent with demand elasticity rather than margin destruction. OpenAI's ~80% cut on GPT-5.6 Luna in July 2026 produced roughly a tenfold usage increase (per CFO comments), helping it gain OpenRouter market share.[19][20] Subsequent September 2026 launches (e.g., GPT-6 Sol/Luna at ~50% lower pricing than predecessors; Anthropic Opus 5.5 with ~20–40% effective savings via efficiency) coincided with further usage spikes and reports of rising token consumption offsetting per-token declines.[21][22]

Analyst notes (e.g., Citadel Securities) highlight that falling per-token costs are fueling additional usage and lifting aggregate AI spending, echoing Jevons Paradox dynamics where efficiency expands total demand.[22][23] This directly challenges concerns that competition or cuts will erode the economics sustaining hyperscaler and lab investments. For pricing strategy, it underscores the value of volume plays and efficiency gains over pure premium positioning.

Eisman's public stance mixes strong concentration warnings with a mostly-long posture and reluctance to call a full implosion, which invites criticism of inconsistency or hedging without full conviction. He has described trimming risk via partial "against the box" shorts on four unnamed AI holdings while remaining "mostly long" and calling a major downturn call "premature."[1][24] His track record is anchored in the successful 2008 subprime short (detailed in The Big Short), with other past commentary including skepticism toward for-profit education and certain financial innovations, but recent coverage emphasizes nuance on bubbles rather than repeated major misses or hits outside that era.[25]

The hedged approach—vocal risks alongside continued long exposure and statements like "I just know I sleep fine"—could be critiqued as allowing participation in upside while positioning for downside without committing to outright shorts, or as inconsistent with the severity of the concentration thesis.[1]

Eisman's thesis would likely fail (or require significant revision) under conditions such as sustained revenue diversification at hyperscalers beyond OpenAI/Anthropic, continued or accelerating token/revenue growth post-price cuts without margin collapse, durable hyperscaler cloud margins and backlog conversion proving capex returns, broad enterprise scaling delivering measurable firm-level ROI, or successful lab IPOs/fundraises demonstrating market confidence in the ecosystem. Key indicators to watch include quarterly hyperscaler cloud growth rates and operating margins (especially Google Cloud and Azure), post-cut usage and revenue elasticity metrics from labs, customer concentration disclosures or diversification trends in RPO/backlogs, enterprise agentic deployment depth and governance maturity from surveys/filings, and actual profitability trajectories or cash-flow inflection in OpenAI/Anthropic updates or prospectuses.[26]

Overall, post-April 2026 developments tilt toward resilience via growth, monetization, and elasticity, though concentration and execution risks remain material variables.