AI在GDP统计中的位置在哪里？（英）2026

Filling the Measurement Gap. Anton Korinek and Patrick McKelvey.May 2026. Anton Korinekhas been anonresident senior fellowat the Peterson Institute forInternational Economics(PIIE) since February 2026and head of TransformativeAI Economic Studies at theAnthropic Institute sinceMay 5, 2026. He is currentlyon educational leave fromthe University of Virginia,where he is professor ofeconomics and facultydirector of the Economicsof Transformative AI(EconTAI) Initiative. Thiswork was conducted in hiscapacity as a nonresidentsenior fellow at PIIE andprofessor at the Universityof Virginia. The artificial intelligence (AI) economy in the United States isgrowing at extraordinary rates of over 2,000 percent per year yetis leaving only a small mark in the nation’s GDP figures. This is ameasurement gap that, left unaddressed, will become a policygap—because what cannot be measured cannot be steered. We estimate in a companionPIIE Working Paper (Korinek andMcKelvey 2026)that nominal AI compute spending grew by morethan 140 percent per year each in 2024 and 2025, raw computecapacity by more than 200 percent per year, and quality-adjustedAI output by more than 2,000 percent per year. The divergencebetween this picture of the AI economy and the one drawn byconventional GDP statistics is itself an informative macroeconomicsignal. Treating the AI sector as a coherent economic entity inits own right yields a preliminary estimate of nominal AI GDPof roughly $250 billion in 2025—comparable in size to the USscheduled passenger airline industry (Bureau of TransportationStatistics 2025)—yet growing at approximately 2,600 percent peryear in quality-adjusted terms. Patrick McKelveyis asenior data scientist at theBank of Canada. The authors thank KodyKarmody and Dylan Ryfefor excellent researchassistance, Leopold Brownand Yuval Rhymon for theircontributions to early-stage research and datacollection, Future ImpactGroup for their support,and Andrey Fradkin forgenerously sharing data oninference prices. We argue that US statistical agencies and economicpolicymakers should start now to assemble better data on AIactivity in AI satellite accounts—focused subsets of the nationaleconomic statistical accounts—in coordination with internationalcounterparts, industry and researchers. They should begin to incorporate AI productive-capacity measures into medium-termprojections and scenario analysis. Building this measurementinfrastructure today, while the AI sector is still small in nominalterms, is needed preparation for a potential phase change in thefuture after which policymakers may need a strong statisticalapparatus to make well-informed decisions about the fast-growing AI economy. TWO PICTURES OF THE SAME ECONOMY. The question of where AI lies in the US GDP statistics hasbecome a recurring mystery in economic commentary. FrontierAI capabilities are advancing at what industry observers considera remarkable pace, with some seeing the possibility of artificialgeneral intelligence within just a few years. Yet when one looksat the conventional national statistical accounts, the AI revolutionregisters only as upstream investment—a data-center capitalinvestment boom—while the productive activity those data centerssupport remains nearly invisible. Overall US GDP growth remainsmoderate, productivity statistics have barely ticked up, and thedisconnect between widely reported AI capability gains and theirmacroeconomic effects has become a source of puzzlement. One natural explanation is that AI adoption takes time, andthe kinds of broad productivity gains economists associate withgeneral purpose technologies typically arrive years after thetechnology itself does. This is almost certainly part of what is goingon. But there is a second explanation, complementary rather thancompeting, that has received much less attention: national accountsas currently constructed have a hard time seeing the AI sector at all. The visibility problem is structural. The conceptual architectureof GDP measurement was developed in the mid-20th centuryto track an economy organized around manufacturing. Thatarchitecture has served well for a long time and continues to do sofor the bulk of economic activity. But it presumes that the structureof the economy changes only slowly, and that quality improvementsin any given sector unfold at a pace that statistical agencies caneasily capture. AI strains that presumption. Quality is improvingso quickly that conventional hedonic adjustment methods,designed for sectors where quality improvements occur at a more moderate pace, do not capture what is happening. Moreover,AI-related activity is dispersed across a long list of industries—cloud services, software publishing, data processing, professionalservices, and others—with no single category that captures the AIeconomy as a whole. AI is the latest in a series of fast-moving technologies that haveraised measurement concerns; semiconductors and the internetgenerated similar deb