Two customers, one AI trade: Eisman says OpenAI and Anthropic are 70% of hyperscaler AI revenue

Steve Eisman told CNBC's "Fast Money" that OpenAI and Anthropic together account for roughly 70% of the AI-related revenue flowing to the major hyperscalers, naming Microsoft, Amazon, Alphabet and Oracle ^[1]. If that number is close to right, the demand story behind hundreds of billions of dollars of infrastructure spending rests on the purchasing decisions of two private AI labs rather than on broad enterprise adoption ^{[5] [12]}.

The arithmetic is the interesting part. Eisman also put AI-driven revenue at approximately 25-35% of those companies' total cloud revenue ^[3]. Applying his own 70% to his own range, two customers would represent something like 17.5% to 24.5% of total cloud revenue at the four firms ^[11]. The residual is more striking: every other AI buyer on earth, every bank, retailer, insurer and government department, would together account for 7.5% to 10.5% ^[13]. That is the entire measured contribution of enterprise adoption in Eisman's framing.

This matters for multiples rather than for quarterly earnings. Cloud has been the growth line that justified premium valuations across the sector for years ^[14], and the capital expenditure case has been built on the assumption that AI workloads earn healthy returns over time ^[15]. Concentration of that kind converts a secular adoption story into a counterparty story: the return on the buildout depends, to a remarkable degree, on the continued spending and success of two companies that do not report to public shareholders ^{[5] [12]}. Eisman calls it the "Achilles' heel" of the AI trade ^[2].

His second point is pricing. Eisman flagged cheaper Chinese open-weight models as a possible trigger for an industry price war ^[6]. Open-weight models publish their parameters for others to build on, and have been gaining traction as alternatives to the proprietary systems from OpenAI and Anthropic ^[7]; Chinese labs have been particularly aggressive on price ^[8]. The damage does not require mass switching. Customers only need a credible alternative to cite in a renewal negotiation for margin assumptions to move.

Two caveats belong in the same breath as the thesis. Eisman is talking his book: earlier in July 2026 he flagged diminishing returns in AI and said he had personally reduced his exposure ^[9]. And the report of his remarks does not disclose where the 70% or the 25-35% comes from, or how "AI-related revenue" is defined ^[16]. He is the investor from "The Big Short" who bet against subprime mortgages ^[4], which buys attention, not verification.

What to watch is whether anyone independently reproduces the 70% figure, because the claim is currently a single sourced estimate carrying a lot of weight. Watch, too, whether the capex announcements keep climbing quarter after quarter as they have through 2026 ^[10] while the customer list stays this short, and what price points the next round of open-weight releases sets ^[8]. A concentration problem and a pricing problem are survivable separately. Arriving together is the scenario worth modelling.