Three years ago, Sequoia partner David Cahn was one of the first to crunch the numbers and calculate the impact of Silicon Valley’s massive spending on AI infrastructure.
In 2023, he was reacting to Nvidia’s reported annual GPU revenue of $50 billion. Starting from this number, and adding in the implicit costs of running a data center and the operator’s margin, he estimated that $200 billion in revenue would be needed to repay the upfront investment.
He took this as a challenge and asked entrepreneurs to come up with AI products and services that leverage all this infrastructure to generate revenue. Fast forward to today and three years of hyperscaling add up and Khan has a new number for AI infrastructure spending in 2026: $1.5 trillion.
All told, he calculates that the AI industry would need to make $3 trillion to justify all the spending on these chips and other data centers. And this is probably an underestimate. Rising costs of memory and increased use of specialized and inference-only chips will further increase this number. “Recently, these bottleneck movements and rising construction costs have sharply increased the required return per GW of capital investment,” he wrote.
Behind the books, Anthropic’s ARR is believed to have reached $60 billion, while OpenAI reportedly made $13 billion in revenue in 2025 (although in November 2025 it announced ARR of $20 billion) and will likely earn even more this year. But it is clear that there is a huge gap to fill.
This gap is of concern to Torsten Slok, chief economist at the giant asset management firm Apollo. In a recent note, he points out that hyperscalers like Google, Meta, Microsoft, and Amazon are all projecting a significant acceleration in free cash flow in 2028. That means they expect to get back on every chip they buy.
What happens if we don’t? Slok points to the risks currently being seen across the use of AI. More organizations are turning to cheaper open-weight models made in China rather than the one built by Frontier Labs, leading to a decline in overall token prices. According to CEO Sam Altman, OpenAI’s latest model is 54% more token efficient on coding tasks. While this is good for users concerned about the cost of AI agents, it could be bad for companies building token factories unless users significantly increase their overall token usage.
Slok worries that if hyperscalers fail to meet cash flow targets, the market reaction could be harsh.
“With so much riding on a small number of stocks, a slowdown in earnings would not just be a sector issue, it would risk sending the economy into recession and the S&P 500 into a correction,” he wrote.
There are some things to keep in mind when directing your AI agent to cheaper tokens.
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
