The AI Bubble's Math Problem: $2T+ Annual Revenue Needed by 2030

Ed Zitron lays out the financial impossibility underlying the current generative AI boom. Roughly 190GW of planned data center capacity, at Jensen Huang’s stated $80–100 billion per gigawatt, implies a $9.5–15 trillion buildout — far above the $3 trillion figure commonly cited. Funding that requires banks to roughly quadruple annual data center debt issuance, while hyperscalers like Google and Meta are increasingly turning to equity sales, a signal that debt markets are tightening.

The demand side is even thinner. OpenAI has made over $770 billion in compute commitments and projects $852 billion in burn through 2030; Anthropic has committed $375 billion across Google, Amazon, Microsoft, CoreWeave, and SpaceX. Together they account for 70–90% of all AI compute demand, yet both lose enormous sums and would need to hit roughly $10 billion in monthly revenue by early 2028 to justify their obligations. Outside these two, Zitron can’t find any buyer spending more than a few hundred million.

The argument isn’t a judgment call about model quality — it’s that NVIDIA’s valuation, hyperscaler capex, and the entire AI infrastructure thesis require generative AI to produce $2 trillion or more in annual revenue by 2030. There is no realistic path to that number from current run rates, and any slowdown collapses the debt-funded chain holding it up.