Are We Building Too Many AI Data Centers?
I am a

Compute bill vs. what AI revenue carries, $ billions a year

Compute bill against what AI revenue carries, 2026 to 2030Two lines in billions of dollars a year. The bill the five buyers pay for compute, and the share of their AI revenue that can carry it. The year the bill rises above that share is marked.

Bill Revenue carries

Tap or hover a year to read it.

Why the crossing matters: The orange line is what the five pay for compute. The blue line is the most their AI revenue can pay for it, after everything else a company has to cover. While the bill is under that line, the buyers have revenue they cannot spend because the compute is not built yet, so they bid against each other for what exists. That is a shortage, and whoever holds compute names the price. Once the bill is over the line, the compute costs more than it earns, and somebody outside, investors, lenders, vendors, pays the difference every year. That is abundance, and the owners cut price to fill the fleet while the users get it cheap. The year the orange line crosses above the blue is the year the market changes hands, and every answer on this page is read from it. More on the year that flips.

What to do

    If you are someone else

    How these answers are decided.

    More
    Defaults are the paper's position.

    From The AI Funding Loop, Han Lee with Claude, November 2026. The five buyers are Anthropic, OpenAI, xAI, Google and Meta. Inputs come from their filings and plans as of 4 November 2026. How the model works · Inputs and sources.