Venture capitalist Tomasz Tunguz has built a model estimating that US data center capacity will expand from 25 gigawatts to 70 gigawatts over the next five years, part of a roughly $5 trillion global buildout financed mostly with debt. The stakes are the credit markets absorbing that debt, and whether AI revenue can grow fast enough to pay it back. Tunguz’s own arithmetic says the growth rate required is steep enough to function as a constraint on the industry, not a comfortable projection.
By Tunguz’s estimate, the new AI-related debt equals roughly a 34 percent expansion of the entire US corporate bond market. That figure alone reframes the buildout: this is not primarily a venture-funded or even an earnings-funded expansion, it is a credit event sized against national bond markets. Tunguz also raises the open question of whether municipalities looking for local economic growth start issuing municipal bonds to help finance individual data center projects, the way they have historically financed power plants and other public infrastructure.
Servicing debt of that scale requires cash flow, and this is where Tunguz’s model gets specific. He calculates that annual AI revenue across software, tokens, and enterprise automation would need to climb from an estimated $150 billion today to somewhere between $1.2 trillion and $1.5 trillion by 2030. Getting from $150 billion to roughly $1.35 trillion in five years works out to a compound annual growth rate near 55 percent.
That number should be read as a derived requirement under a specific set of assumptions, not as a forecast anyone has committed to. Tunguz builds it from an assumed interest rate band, an assumed interest coverage ratio, and an assumed gross margin on AI infrastructure revenue. Change any one of those inputs and the required growth rate moves substantially: a lower interest rate, a higher coverage cushion lenders are willing to accept, or fatter margins would all reduce the 55 percent figure, while a rougher credit market or thinner margins would push it higher. The model is useful precisely because it makes the assumptions visible, not because the output is a settled number.
The genuinely interesting part of a debt-financed buildout is structural rather than statistical. A bond payment schedule does not adjust itself if enterprise AI adoption arrives two years late. That turns the growth rate from a nice-to-have projection into a hard constraint: the industry either hits something close to that revenue trajectory or the financing structure underneath the buildout comes under stress regardless of how promising the underlying technology looks on a slower timeline.
Tunguz does offer one comparison point that cuts against pure skepticism: current hyperscaler cloud growth rates already range from 37 percent at AWS to 82 percent at Google Cloud, with Azure in between at 43 percent. A 55 percent blended growth rate for AI-specific revenue sits inside that existing range rather than far outside it, though it would require the slower-growing providers to accelerate meaningfully or a shift in the revenue mix toward the fastest-growing categories.
Operators building on infrastructure financed by this debt wave should treat the 55 percent CAGR as a floor to watch, not a target to assume. If quarterly hyperscaler disclosures show AI infrastructure revenue growth drifting below that pace for two or three consecutive quarters, it becomes a leading indicator of credit stress in the data center financing market well before any single default makes headlines.
Tomasz Tunguz laid out this debt-financing model in a post on his own site, tomtunguz.com.
