Can AI answer the $3 trillion question?

Three years ago, Sequoia partner David Cahn was among the first to do the calculations and provide numbers on the implications of Silicon Valley’s massive spending on AI infrastructure.

In 2023he responded to Nvidia’s reported annual GPU revenue of $50 billion. Taking that figure, and adding in the implicit costs of running the data centers and the margins for their operators, he concluded that $200 billion in revenues would be needed to repay the initial investment.

He took it as a challenge and asked entrepreneurs to come up with AI products and services to leverage and monetize all that infrastructure. Fast forward to today, three years of hyperscaling combined, and Cahn has a new number on AI infrastructure spending for 2026: $1.5 trillion.

All told, he calculates that the AI ​​industry will need to earn $3 trillion to justify all those chips and other data center expenses. And that’s probably an underestimate: rising memory costs and the increasing use of exotic or inference-specific chips will increase that number. “Recently,” he writes, “required revenue per GW of CapEx has risen sharply as a result of these bottleneck dynamics and rising construction costs.”

On the other side of the ledger, Anthropic is believed to have struck $60 billion in ARRwhile OpenAI reportedly earned $13 billion in 2025 (although in November 2025 it said ARR was $20 billion) and is likely to earn more this year. But there is clearly a big gap that needs to be closed.

One person paying attention to this gap is Torsten Slok, the chief economist at Apollo, the giant asset manager. In one recent noteHe points out that the hyperscalers – Google, Meta, Microsoft and Amazon – are all predicting massive accelerations in their free cash flow by 2028. That means they expect the payback on all those chips they bought.

What if they don’t? Slok notes a risk we currently see in using AI: more and more organizations are turning to cheaper open-weight models, often Chinese ones, not the ones built by the frontier labs, and overall token prices are falling. OpenAI’s latest model, according to CEO Sam Altman, is 54% more token efficient about coding tasks. That’s good for users concerned about the cost of their AI agents, but it could be bad for companies building token factories if users don’t wildly increase their overall token usage.

Slok worries that if hyperscalers don’t meet their cash flow targets, the market reaction could be severe:
“Because so much depends on so few names,” he writes, “a slower payout would not only be an industry problem, but would also risk sending the economy into a recession and the S&P 500 into a correction.”

Something to keep in mind as you guide your AI agents to cheaper tokens.