Investigate Eisman's views on whether AI model and infrastructure companies have durable moats, and his predictions about price wars.

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.