US AI economy hits $250B in 2025, but GDP stats miss it entirely

Economists document a $250 billion AI sector growing at 2,600% annually—that GDP can't see

Researchers at the University of Virginia, Anthropic, and the Bank of Canada published findings showing nominal AI GDP in the United States reached approximately $250 billion in 2025, growing at roughly 2,600 percent per year in quality-adjusted real terms (per the co-authored paper). Yet standard GDP statistics show almost no anomaly.

The measurement gap stems from two structural misalignments. First, most AI economic activity happens in inference—the use of trained models in production—not in the visible capex surge of datacenter buildout. Second, per-unit prices for a given capability fall nearly as fast as algorithmic improvements rise, so nominal revenues grow only moderately even as underlying productive capacity accelerates.

Three ways to measure the gap illustrate the scale:

• Nominal compute spending: US compute spending rose from $37 billion in 2023 to $90 billion in 2024 to $219 billion in 2025.
• Raw compute capacity: Due to chip efficiency gains, actual capacity grew at more than 200 percent per year.
• Quality-adjusted output: Factoring in algorithmic progress and inference price declines, quality-adjusted AI output grew at roughly 2,290 percent in 2024 and 2,271 percent in 2025.

The authors note that conventional statistics show a sector growing slowly in nominal terms, while their measures show one whose underlying capacity more than doubles annually.

Labor displacement was always the threat; mismeasurement is now the policy risk

Prior fast-moving technologies—semiconductors, the internet—raised similar measurement debates. But AI differs in one critical way: those technologies were complements to human labor at the aggregate level. AI is the first plausible candidate for large-scale technological mismeasurement where the rapidly improving sector may become a substitute for human labor.

A finance ministry running ten-year revenue projections off conventional GDP data will materially underweight the probability of a labor-tax-base shock. That undercounting leaves it unprepared to design responses such as tax system reforms, sovereign wealth funds, or benefit-sharing schemes that such a shock would call for.

The gap also creates what the authors call a "windfall that cannot be seen cannot be shared." If policymakers don't perceive the true shape of AI's economic contribution, the gains concentrate in whoever owns the compute and models, while society bears the labor costs.

Three levers to close the measurement gap

The authors propose three interventions for statistical agencies and policymakers:

• AI satellite accounts: Statistical agencies should develop measures (e.g., nominal compute spending, training versus inference allocation) that can inform overall GDP calculations without waiting for conventional accounting to adapt.
• Better primary data: Partner statistical agencies, companies, and academia to generate shared datasets on compute allocation, model training costs, and inference usage patterns.
• Medium-term projections: Policymakers should incorporate AI productive-capacity measurements into fiscal and labor-market planning, not just wait for GDP to catch up.

The practical implication is stark: the intuitions of everyone building or working in AI are that the economy is in the midst of a discontinuity. Conventional economic data says otherwise. One group is right. One group is blind. The gap between them is where policy surprises live.