Sequoia's Cahn Pegs AI's 2026 Revenue Gap at Roughly $3T

The math that Sequoia partner David Cahn has been running on AI infrastructure for three years now has quietly become the single scariest number in the industry. In his latest update, reported by TechCrunch, Cahn puts 2026 AI infrastructure spending at $1.5 trillion and calculates the industry needs to earn roughly $3 trillion in revenue to justify it. For context, back in 2023 he was reacting to Nvidia's $50 billion in annual GPU revenue and pegged the payback gap at $200 billion. The hole grew fifteen-fold in three years.

The revenue side is where it gets uncomfortable. Anthropic is reportedly near $60 billion in annual recurring revenue and OpenAI hit a $20 billion annualized run rate by November 2025, having booked $13 billion in 2025 revenue. That is real business, but combined it barely dents a $3 trillion target. Cahn's own framing is that "the required revenue per GW of CapEx has sharply increased due to these bottleneck dynamics and rising costs of construction," meaning the goalposts are moving away from the frontier labs, not toward them.

Why this reads as a macro conversation rather than a Sequoia blog post: Apollo chief economist Torsten Slok points out that Google, Meta, Microsoft and Amazon are all projecting large free cash flow accelerations by 2028, and with the index this concentrated in those four names, "a slower payoff wouldn't just be a sector problem, it would risk tipping the economy into recession and the S&P 500 into a correction." That is the tail attached to the capex bet.

The honest caveat is that Cahn's exercise is a top-down payback model, not a demand forecast, and it doesn't price in the enterprise adoption curve that is still early. What the reporting doesn't give you is utilization rates on the capacity already built, or how much of the $1.5 trillion is contractually committed versus optionality hyperscalers can pace. It also flags a headwind cutting the other way: OpenAI's newest model is reportedly 54% more token-efficient on coding tasks than its predecessor, which is great for buyers and awkward for anyone modelling revenue as tokens sold.

The forward-looking read is that the pressure is going to land unevenly. If open-weight substitution and token deflation keep compounding, the surplus flows to the application layer and to the buyers, while the exposure sits with whoever is holding the depreciation schedule on those gigawatts of capacity.