衡量人工智能经济（英）

Anton Korinek and Patrick McKelveyMay 2026 ABSTRACT Anton Korinekhas been anonresident senior fellowat the Peterson Institutefor InternationalEconomics (PIIE) sinceFebruary 2026 and headof Transformative AIEconomic Studies at theAnthropic Institute sinceMay 5, 2026. He is onleave from the Universityof Virginia (UVA),where he is professor ofeconomics and facultydirector of the Economicsof Transformative AI(EconTAI) initiative. Thiswork was conductedin his capacity as PIIEnonresident senior fellowand professor at UVA.Patrick McKelveyis asenior data scientist at theBank of Canada. We construct a macroeconomic estimate of total AI production forthe United States, combining inference and R&D/training activitiesand applying quality adjustments based on the evolution of API pricesat fixed performance levels and the pace of algorithmic progress. Weestimate that nominal AI compute spending grew over 140 percentper year each in 2024 and 2025, raw compute capacity grew over200 percent per year, and quality-adjusted AI output grew over 2,000percent per year. These growth rates reflect three compoundingforces: expanding data-center capacity, continued improvements inchip efficiency, and rapid algorithmic progress. We then employ ourestimates to develop a nascent framework for “AI GDP” that tracks theAI economy as a coherent whole rather than dispersed across standardindustry classifications. Quality-adjusted AI GDP grew by more than2,500 percent each in 2024 and 2025. Our measures complementtraditional national accounts by providing visibility into a fast-movingsector whose activity is difficult to isolate in existing statistics, andthey may serve as building blocks for satellite accounts that track AI’sgrowing role in the economy. The authors thank MartinChorzempa, CullenHendrix, Patrick Honohan,Adam Posen, and DavidWilcox for excellentcomments; Kody Karmodyand Dylan Ryfe for reliableresearch assistance;Leopold Brown andYuval Rhymon for theircontributions to early-stage research and datacollection; Future ImpactGroup for their support;and Andrey Fradkin forgenerously sharing dataon inference prices. JEL Codes:E01, O33, O47, E22 Keywords:artificial intelligence; national accounts; GDP mismeasure-ment; AI satellite accounts; quality-adjusted prices; algorithmic prog-ress; AI GDP The findings, interpretations, views, and conclusions expressed herein aresolely those of the authors and do not necessarily represent those of theBank of Canada, the Anthropic Institute, or the Peterson Institute forInternational Economics. The authors have used AI extensively at everystage of the research and writing process and have subjected all AI-generated output to careful human review. Contents 1Introduction32ApproachandContributions43Methodology73.1Measuring AI Production. . . . . . . . . . . . . . . . . . . . . . . . . .74Results104.1Quality-Adjusted AI Production. . . . . . . . . . . . . . . . . . . . . . .115FromAIProductiontoAIGDP116Discussion157Conclusion18 1Introduction Amongartifcialintelligence(AI)researchersandleadingtechnologycompanies,thereisbroadagreementthatAIcapabilitiesareadvancingataremarkablepace—withsomearguingthatartifcialgeneralintelligence(AGI)maybeachievedsoon.Yetwhenwelookattraditionaleconomicstatistics,weseeonlyupstreaminvestmentindatacenters,whiledownstreamimpactsfromthisrevolutionremainnearlyinvisible.GDPgrowthintheUnitedStatesandotheradvancedeconomieshasremainedmoderate,andproductivitystatisticshavebarelytickedup.Thequestion“whenwillweseeAIintheGDPstatistics?”hasbecomearecurringthemeineconomiccommentary.Onenaturalresponseispatience:AIadoptiontakestime,andtransformativeeconomicefectsmaysimplylieahead.Thisisalmostcertainlypartofthestory. Butwebelievethereisanadditional,complementaryissueworthtakingseriously.Nationalaccountsweredesignedforaneconomyinwhichallproductionisultimatelyorganizedaroundhumansasthecentralpointofvaluecreation.Thiswasanentirelyappropriatedesignformostofeconomichistory,anditcontinuestoserveitscorepurposewell.However,therapidgrowthoftheAIsectorintroducesmeasurementchallengesthatexistingstatisticalcategorieswerenotbuilttoaddress.ThedifcultyisthatAIactivityishardtoseethroughthelensoftraditionalnationalaccountsandinthewayswetypicallymeasureGDP. Thechallengeoperates throughseveral channels.First,AIactivityis scattered:spendingonAIcompute,modeldevelopment,andAI-poweredservicesisspreadacrossdozensofindustrycategories—dataprocessing,cloudcomputing,softwarepublishing,professionalservices—makingitdifculttotracktheAIeconomyasacoherentwhole.Second,AIqualityimprovementsareunusuallyrapid:thepaceofimprovementinAIcapabilitiesisfarfasterthaninmostsectorsforwhichstatisticalagencieshavedevelopedqualityadjustmentmethods,raisingquestionsaboutwhetherstandardhedonictech-niquescapturewhatishappening.Third,AI’sroleintheeconomyisevolving:asAIsystemsbecomemorecapable,theymaytransitionfrombeingoneamongmanyinterme-diateinputstoplayingamorecentralroleinproduction,potentiallystrainingcategoriesthatweredesignedf