The US AI economy is growing at roughly 2,000% a year and almost none of it shows up in headline GDP, according to a new paper from economists at the University of Virginia, Anthropic, and the Bank of Canada. The authors put nominal AI GDP at around $250 billion in 2025 and quality-adjusted real growth at about 2,600% per year. Conventional statistics, they argue, are mismeasuring a sector whose underlying capacity is more than doubling annually.
The numbers underneath the headline are unusual on their own terms. US compute spending rose from $37 billion in 2023 to $90 billion in 2024 to $219 billion in 2025 — a near-sixfold jump in two years. Raw AI computing capacity grew faster still, at more than 200% per year, because newer chips deliver more useful work per dollar. Layer in algorithmic efficiency gains and the authors estimate quality-adjusted AI output grew 2,290% in 2024 and 2,271% in 2025.
“Treating the AI sector as a coherent economic entity yields preliminary estimates of nominal AI GDP at approximately $250 billion in 2025, growing at roughly 2,600 percent per year in quality-adjusted real terms.”— Anton Korinek, University of Virginia economist
Nominal revenue tells a much tamer story, and the paper explains why. Per-unit prices for any given level of AI capability are falling almost as fast as quality-adjusted output is rising, so the dollars booked by AI providers grow only moderately even as the real economic surface area explodes. That decoupling is exactly what makes the sector invisible to standard GDP accounting, which tracks nominal spending rather than capability delivered.
Key facts
- 01Nominal AI GDP in the US hit roughly $250 billion in 2025, growing at 2,600% per year in quality-adjusted real terms.
- 02US compute spending climbed from $37 billion in 2023 to $90 billion in 2024 to $219 billion in 2025.
- 03Quality-adjusted AI output grew an estimated 2,290% in 2024 and 2,271% in 2025, with US AI computing capacity doubling annually.
- 04The paper was co-authored by Anton Korinek of the University of Virginia, with researchers from Anthropic and the Bank of Canada.
- 05The authors call for AI satellite accounts in national GDP statistics to make labor-tax-base risk visible to finance ministries.
The authors compare the moment to earlier mismeasured booms in semiconductors and the internet, when statistical agencies took years to catch up to what the underlying technology was actually doing. The disanalogy they flag is the one that matters most for policy. Semiconductors and the internet complemented human labor in aggregate; AI is the first such technology with a credible path to substituting for it at scale.
That framing is the policy punchline. A finance ministry running ten-year revenue projections off conventional data, the paper warns, will materially underweight the probability of a labor-tax-base shock and will be unprepared to design responses — tax reform, sovereign wealth funds, or other benefit-sharing schemes — that such a shock might require. "A windfall that cannot be seen cannot be shared," the authors write.
The recommendations are concrete. Statistical agencies should build dedicated AI satellite accounts that track nominal compute spending, the training-versus-inference split, and quality-adjusted output as standing series. Agencies, companies, and academic researchers should partner to generate the primary data those accounts need. And medium-term economic projections from policymakers should incorporate AI productive-capacity measurements rather than extrapolating from historical labor-share trends.
“AI is the first plausible candidate for large-scale technological mismeasurement in which the rapidly improving sector may become a substitute for human labor.”— Anton Korinek, University of Virginia economist
Anton Korinek, the University of Virginia economist on the paper, is also affiliated with Anthropic, though the bulk of the research was completed before he joined the company. The Bank of Canada coauthorship signals that central banks are starting to take the measurement gap seriously, not just AI labs with an interest in being seen as economically consequential.
The caveats matter. Quality-adjusted output numbers depend heavily on assumptions about how to price algorithmic progress and inference-cost declines, and the 2,600% headline figure sits inside a wide confidence band. Skeptics will note that capacity doubling annually is not the same as economic value doubling annually — much of the new compute is being burned on training runs whose commercial return remains unproven. The paper's own framing concedes that conventional nominal measures are not wrong, just incomplete.
Still, the gap between what people working inside AI see day-to-day and what aggregate economic statistics show is now wide enough that serious economists are putting numbers on it. If the authors are even directionally right, the labor-market and fiscal implications arrive faster than the data series tracking them. That asymmetry — capability moving faster than measurement — is the actual policy problem.
For AI companies, the paper is useful ammunition in conversations with regulators and finance ministries that have been slow to treat the sector as macroeconomically material. For investors, it reframes the compute buildout — $219 billion in 2025 US spending alone — as the visible tip of a much larger productive base. And for anyone modeling the next decade of labor demand, it argues that the conventional dashboards are reading the wrong dial.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
