Tomasz Tunguz argues that AI data center construction now depends on the global credit markets. He estimates hyperscalers and data center operators must issue about $4t of debt over five years. Data centers are financed like real estate projects, with debt usually above 70% of the cost. He compares that borrowing to the size of the commercial paper, private credit, corporate bond and municipal bond markets. He also asks whether local governments will issue municipal bonds to fund such projects, as they do for power plants. Servicing the debt would require annual AI revenue far above today's estimated level. That implies growth faster than any large cloud provider currently reports. The figures come from a post by Tunguz, a partner at Theory Ventures, drawing on research from J.P. Morgan, PIMCO, Gartner and others.

• $4t of AI data center debt in 5 years
• US capacity 25GW to 70GW
• 286% of US commercial paper
• 55% CAGR needed to $1.35t revenue