Three years in the past, Sequoia accomplice David Cahn was one of many first individuals to do the maths and put a quantity on the implications of Silicon Valley’s titanic spend on AI infrastructure.
In 2023, he was reacting to Nvidia’s reported annual GPU income of $50 billion. Beginning with that determine, and including within the implied prices of working the info facilities and the margins for his or her operators, he deduced that $200 billion in income could be required to pay again the up-front funding.
He took it as a problem, asking entrepreneurs to give you AI services to utilize, and generate income from, all that infrastructure. Quick-forward to right now, including up three years of hyperscaling, and Cahn’s acquired a new number on AI infrastructure spending for 2026: $1.5 trillion.
All instructed, he calculates that the AI trade should earn $3 trillion to justify all these chips and different information heart expenditures. And that’s most likely an underestimate — the rising prices of reminiscence and the growing use of unique or inference-specific chips will drive that quantity up. “Lately,” he writes, “the required income per GW of CapEx has sharply elevated on account of these bottleneck dynamics and rising prices of building.”
On the opposite facet of the ledger, Anthropic is believed to have hit $60 billion in ARR, whereas OpenAI reportedly earned $13 billion in 2025 (though in November 2025, it said it was at $20 billion ARR) and is presumably making extra this 12 months. However there’s clearly a big hole to be closed.
Somebody minding that hole is Torsten Slok, the chief economist at Apollo, the large asset supervisor. In a recent note, he factors out that the hyperscalers — Google, Meta, Microsoft, and Amazon — are all predicting huge accelerations of their free-cash stream in 2028. That’s, they anticipate to see the payback from all these chips they purchased.
What in the event that they don’t? Slok notes a threat we’re at present seeing throughout AI utilization: Extra organizations turning to cheaper open weight fashions, usually Chinese language, not these constructed by the frontier labs, and general token costs falling. OpenAI’s newest mannequin, per CEO Sam Altman, is 54% more token efficient on coding duties. That’s good for customers fretting about the price of their AI brokers, however it could be dangerous for corporations constructing token factories ought to customers not wildly improve their general token utilization with them.
Slok worries that if hyperscalers don’t meet their cash-flow objectives, the market response may very well be extreme —
“with a lot using on so few names,” he writes, “a slower payoff wouldn’t simply be a sector drawback, it might threat tipping the economic system into recession and the S&P 500 right into a correction.”
Simply one thing to remember as you’re herding your AI brokers towards cheaper tokens.
If you buy by means of hyperlinks in our articles, we may earn a small commission. This doesn’t have an effect on our editorial independence.
