Sequoia partner David Cahn now estimates the AI industry will need to generate $3 trillion in revenue to justify the chips and data centers being built through 2026, up from the $200 billion figure he first calculated three years ago. Cahn's new note puts 2026 AI infrastructure spending alone at $1.5 trillion. The two largest frontier labs are nowhere near closing that gap: Anthropic is thought to have hit $60 billion in annualized revenue, while OpenAI earned $13 billion in 2025 and said in November 2025 it had crossed $20 billion ARR.
The math has moved fast. In 2023, Cahn started from Nvidia's reported $50 billion in annual GPU revenue, layered on data-center operating costs and operator margins, and arrived at a $200 billion revenue requirement to pay back the up-front infrastructure investment. Three years of hyperscaler capex later, that same exercise produces a number 15 times larger. Cahn treats the figure as an underestimate, citing rising memory costs and the growing share of exotic or inference-specific silicon in new builds.
Cost per gigawatt of capacity is a big part of the shift.
Key facts
- 01Sequoia's David Cahn pegs 2026 AI infrastructure spending at $1.5 trillion, requiring $3 trillion in revenue to justify the buildout.
- 02Anthropic is thought to have reached $60 billion in ARR; OpenAI earned $13 billion in 2025 and said it hit $20 billion ARR by November 2025.
- 03Cahn's original 2023 model, built off Nvidia's $50 billion in GPU revenue, implied a $200 billion revenue requirement — the gap has grown 15x since.
- 04Google, Meta, Microsoft, and Amazon are all forecasting sharp free-cash-flow acceleration in 2028 as the payoff window for current AI capex.
- 05OpenAI's latest model is 54% more token-efficient on coding tasks per Sam Altman — good for users, bad for token-factory unit economics.
Add the two frontier labs together and the industry is running at roughly $80 billion in annualized revenue against a $3 trillion target. Even generous assumptions about growth from Google, Microsoft, Meta, Amazon, and the long tail of AI-native startups leave a wide crevasse. The bull case is that inference demand compounds fast enough — over enough surface area — that the $3 trillion figure gets absorbed before the depreciation schedules catch up with the P&L.
The bear case has a name: Torsten Slok, chief economist at Apollo. In a recent note, Slok points out that Google, Meta, Microsoft, and Amazon are all guiding investors toward a sharp acceleration in free cash flow in 2028 — the year hyperscalers are implicitly promising the AI capex will start paying back. Every quarterly investor update between now and then is a bet on that curve holding.
Slok flags two forces already pressing on it. The first is that more organizations are running cheaper open-weight models — often Chinese — rather than paying frontier-lab pricing. The second is that token prices are falling across the board. OpenAI CEO Sam Altman says the company's latest model is 54% more token-efficient on coding tasks than its predecessor, a genuine win for developers but a headwind for anyone selling tokens at scale unless usage volumes rise faster than per-token prices fall.
The concentration risk is the part Slok keeps returning to. The four hyperscalers plus Nvidia now account for an outsized share of S&P 500 earnings growth and index-level performance.
There are ways the gap closes. Enterprise AI budgets are still early — most Fortune 500 deployments today are pilots, not full production replacements — and the shift from chatbot pricing to agent pricing (per-task, per-outcome) could reprice the market upward. Inference for reasoning models consumes far more compute per query than the last generation, which is bad for margins but bullish for capacity utilization. And the actual free-cash-flow guidance from the hyperscalers is for 2028, giving the industry two more years to show the demand curve.
There are also ways it doesn't. Token deflation could outrun usage growth. Open-weight models could take a larger share of the workloads that today underwrite frontier-lab revenue. Enterprise deployment timelines could stretch as compliance, evaluation, and integration costs pile up. Any one of these on its own is manageable; two or three at once turn the 2028 cash-flow story into a 2030 cash-flow story, which is a very different conversation with public-market investors.
The $3 trillion number is worth taking seriously not because it's a forecast but because it's a hurdle rate. It sets the size of the AI economy that has to exist by the late 2020s for today's capex decisions to look rational in hindsight. The frontier labs are growing fast enough — Anthropic's climb from a standing start to a reported $60 billion ARR is the fastest revenue ramp in enterprise software history — but the hyperscalers are spending faster still. Whichever side is compounding harder over the next 24 months will decide whether Cahn's gap gets closed or gets repriced by the market on the way down.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
