Ed Zitron: AI Compute Buildout Demands $2T Revenue By 2030

Ed Zitron's latest essay on Where's Your Ed At reads like a stress test of the numbers everyone in the sector has been quietly hoping nobody would total up. He puts the planned data center buildout at 190GW, with a price tag he pegs at $9.5 trillion to $15 trillion, and argues the only way that math works is if generative AI and AI compute generate over $2 trillion in annual revenue by 2030.

To get there, the two companies that matter most have to grow at speeds the cloud business has never seen. Zitron writes that Anthropic has made $330 billion in compute and chip commitments between Google, Amazon and Microsoft, plus another $30 billion with CoreWeave and $15 billion with SpaceX, and that it must meet its projected revenue of $174 billion a year by 2029 for any of that to be solvent. OpenAI, on his accounting, has projected to burn at least $852 billion through the end of 2030, with $50 billion earmarked for compute in 2026 alone. Together the two firms reportedly account for 89% of all AI startup revenues, and Zitron notes that 54% of NVIDIA's revenue currently comes from three clients, so any softening at the top hits the whole chain.

The more interesting part of the piece is not the macro number, it is the demand-side detail. Zitron walks through token caps that companies have started imposing on their staff. Uber has capped employee spend at $1,500 a month per user, with T-Mobile at $2,000 a month, and Brex is limiting engineers to $500 a week in tokens and non-engineers to $5 a week. He cites a survey finding that 26% of companies say they have a comprehensive view of their AI costs, while 22% report no visibility or visibility only after billing. The implication, when CFOs finally see the line items, is that they cut. Microsoft AI's Mustafa Suleyman is quoted saying Anthropic's models were too expensive and that he intended to reduce Microsoft's use of them to zero.

The honest caveat is that this is one columnist's reading of contracts, projections and survey data, and Zitron is openly skeptical of the sector, so take the specifics as reported, not settled. What the piece does not give you is a sober counter-scenario, what real recurring AI workloads beyond chat actually look like at scale, or how much of the $2 trillion target an unbiased demand model would support.

If even the directional math is right, the people who benefit are the buyers of compute, not the sellers. Smaller models, on-prem inference and finance tooling that gives real-time visibility into token spend all get more interesting in a world where hyperscalers have to defend their commitments rather than expand them.