The $80 Billion AI Revenue Mirage: One Customer Dollar May Be Counted by the App, Model and Cloud

The 89% Duopoly Also Reveals an Accounting Maze

The Information tracks 34 companies including OpenAI, Anthropic, Cursor, Cognition and ElevenLabs. Their combined annualized revenue approaches $80B, or $6.6B monthly, after 112% growth in six months. The two labs capture 89%, up 4.5 points.

But the pool combines foundation models with applications that buy those models. If an enterprise pays Cursor $100, Cursor pays Anthropic $40 and Anthropic pays a cloud provider, one end-customer budget produces revenue at three companies.

The public summary notes Anthropic sharing revenue with Amazon and Google resellers, while OpenAI continues Microsoft sharing through 2030 with potential payments capped at $38B. Each layer records real revenue, but the layers are not independent demand.

Who Actually Keeps the $80 Billion?

Mature SaaS often earns 75%–85% gross margins because another customer costs little to serve. Every AI response, code generation or agent task creates inference cost. Heavy users can be less profitable than light ones.

CNBC reported roughly $47B of annualized revenue at Anthropic and $25B at OpenAI. Those figures explain most of the pool but do not reveal cloud sharing, inference cost or acquisition subsidies.

Cursor crossed $1B in annualized revenue with just over 300 employees in late 2025; Perplexity, ElevenLabs and Cognition have crossed $500M. Demand is real. The unanswered question is whether customers pay for an irreplaceable workflow or resold tokens.

A Better Lens: Multi-Layer Recognition, Not Fake Revenue

Every layer provides a real service, so none of the revenue should simply be deleted. Yet adding every layer and calling it end-market demand exaggerates the economic pool.

Separate end AI spend, total value-chain revenue and retained gross-profit pools. The first measures customer budgets. The second measures how often money circulates. The third—what remains after upstream costs—best approximates shareholder value.

This explains how revenue can grow 112% while the industry still requires extraordinary financing. Each growth dollar may carry inference, cloud commitments, sharing and subsidies. Revenue velocity measures adoption; retained gross profit measures economics.

What Buffett, Graham, Munger, Lynch and Marks Would Test

Buffett would follow owner earnings after inference, training and the cost required to maintain capability. Costs needed to keep customers are not freely distributable cash.

Graham would test earnings quality. Annualized figures multiply the latest month by twelve and can overstate durability during exponential growth. He would demand actual quarters, concentration, retention and normalized margins.

Munger would invert: why will customers stay if the model vendor copies the feature, the API rises 30%, or a cheap open model reaches 80% quality?

Peter Lynch would verify daily workflow, embedded data and which existing budget the tool replaces. “Uses AI” is not a story; a displaced cost center is.

Howard Marks would watch the profit pool migrate as scarce models become abundant. An 89% share is today’s structure, not a perpetual law. Paying a perpetual multiple for cyclical scarcity is the danger.

Replace ARR With Retained Gross-Profit Revenue

Determine whether revenue is gross or net. Separate marketplace resale, model pass-through and partner shares. Deduct API, inference cloud, free credits and heavy-user subsidies. Then stress a 30% supplier price increase and a 70% model-price decline: one tests cost risk, the other copy risk.

Measure multi-model portability. Routing across OpenAI, Anthropic, Google and open models creates procurement option value; single-model dependency deserves a higher discount rate. Translate data, workflow and outcome pricing into retention, expansion, task success and customer savings.

In the Valuefocus DCF calculator, model inference as a separate percentage of revenue rather than assuming SaaS margins. Use the margin-of-safety calculator to test whether valuation survives greater supplier capture, and read our AI capex analysis for foundation-model capital needs.

The 89% figure is neither an application-layer death certificate nor permanent proof of a model duopoly. It is an X-ray of revenue circulating rapidly through one value chain. The moat belongs to the company that keeps each customer dollar longest and turns the most of it into free cash flow.