A working paper’s central scenario puts the annual revenue needed to support the planned US AI data-centre build-out at about $3.55 trillion once the facilities are mature. The eye-catching figure is a modelled requirement, not a prediction that AI companies will actually earn that much.
Watch Desk analysis
What happened
The estimate comes from a working paper by Columbia Business School finance professor Stijn Van Nieuwerburgh, as described by The AI Insider in a report published on 9 October. Its central scenario assumes about 188 gigawatts of additional US data-centre capacity comes online during 2025–2032, with nearly $9 trillion invested over the period. Read The AI Insider’s report.
The calculation assumes a 50 per cent operating cash-flow margin and a six-year economic life for computing equipment. Under those assumptions, the facilities would need about $1.77 trillion in annual operating cash flow, equivalent to roughly $3.55 trillion in revenue. If the equipment’s assumed economic life falls to three years, the report says, the revenue requirement rises to about $5.7 trillion.
Why it matters
This is a way of asking what the infrastructure boom must earn to support its investment, not a sales target for any single AI company or a measure of AI’s wider economic value. The answer depends heavily on assumptions about financing costs, profit margins, equipment lifespans, project cancellations and how much capacity gets used.
Utilisation matters too. The report says the model’s required revenue translates to about $5.10 per installed GPU-hour at full utilisation, rising to $6.40 at 80 per cent and $7.30 at 70 per cent. Those are modelled revenue requirements, not forecasts of rental prices.
Our read
The useful point is not that the build-out must fail, or that these figures are a bill someone has already received. It is that the economics are sensitive to how long expensive equipment stays competitive and how often it earns money. “The AI boom” is not one tidy balance sheet; this paper gives readers a concrete set of assumptions to challenge.
What to watch
- Whether completed capacity and investment track the paper’s central scenario.
- How quickly newer hardware makes existing equipment less attractive to customers.
- Whether providers can sustain high utilisation without prices falling.
- How future research accounts for leases, joint ventures and borrowing outside technology-company balance sheets.
Discussion spark: How much should investors and policymakers trust a model whose result shifts so sharply with equipment lifespan?
Sources and evidence
- Economist Estimates AI Buildout Would Need $3.55 Trillion in Annual Revenue (9 October 2026, 14:30 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.