版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领
文档简介
WORKINGPAPER
26-9MeasuringtheAIEconomy
AntonKorinekandPatrickMcKelvey
May2026
ABSTRACT
AntonKorinekhasbeenanonresidentseniorfellow
WeconstructamacroeconomicestimateoftotalAIproductionfor
atthePetersonInstitute
theUnitedStates,combininginferenceandR&D/trainingactivities
forInternational
andapplyingqualityadjustmentsbasedontheevolutionofAPIprices
Economics(PIIE)sinceFebruary2026andhead
atfixedperformancelevelsandthepaceofalgorithmicprogress.We
ofTransformativeAI
estimatethatnominalAIcomputespendinggrewover140percent
EconomicStudiesatthe
peryeareachin2024and2025,rawcomputecapacitygrewover
AnthropicInstitutesinceMay5,2026.Heison
200percentperyear,andquality-adjustedAIoutputgrewover2,000
leavefromtheUniversity
percentperyear.Thesegrowthratesreflectthreecompounding
ofVirginia(UVA),
forces:expandingdata-centercapacity,continuedimprovementsin
whereheisprofessorofeconomicsandfaculty
chipefficiency,andrapidalgorithmicprogress.Wethenemployour
directoroftheEconomics
estimatestodevelopanascentframeworkfor“AIGDP”thattracksthe
ofTransformativeAI
AIeconomyasacoherentwholeratherthandispersedacrossstandard
(EconTAI)initiative.Thisworkwasconducted
industryclassifications.Quality-adjustedAIGDPgrewbymorethan
inhiscapacityasPIIE
2,500percenteachin2024and2025.Ourmeasurescomplement
nonresidentseniorfellow
traditionalnationalaccountsbyprovidingvisibilityintoafast-moving
andprofessoratUVA.PatrickMcKelveyisa
sectorwhoseactivityisdifficulttoisolateinexistingstatistics,and
seniordatascientistatthe
theymayserveasbuildingblocksforsatelliteaccountsthattrackAI’s
BankofCanada.
growingroleintheeconomy.
TheauthorsthankMartinChorzempa,Cullen
Hendrix,PatrickHonohan,
JELCodes:E01,O33,O47,E22
AdamPosen,andDavid
Keywords:artificialintelligence;nationalaccounts;GDPmismeasure-
Wilcoxforexcellent
ment;AIsatelliteaccounts;quality-adjustedprices;algorithmicprog-
comments;KodyKarmodyandDylanRyfeforreliable
ress;AIGDP
researchassistance;
LeopoldBrownand
YuvalRhymonfortheir
Thefindings,interpretations,views,andconclusionsexpressedhereinare
contributionstoearly-
solelythoseoftheauthorsanddonotnecessarilyrepresentthoseofthe
stageresearchanddata
BankofCanada,theAnthropicInstitute,orthePetersonInstitutefor
collection;FutureImpact
InternationalEconomics.TheauthorshaveusedAIextensivelyatevery
Groupfortheirsupport;
stageoftheresearchandwritingprocessandhavesubjectedallAI-
andAndreyFradkinfor
generatedoutputtocarefulhumanreview.
generouslysharingdataoninferenceprices.
2
Contents
1Introduction
3
2ApproachandContributions
4
3Methodology
7
3.1MeasuringAIProduction
7
4Results
10
4.1Quality-AdjustedAIProduction
11
5FromAIProductiontoAIGDP
11
6Discussion
15
7Conclusion
18
3
1Introduction
Amongartificialintelligence(AI)researchersandleadingtechnologycompanies,thereisbroadagreementthatAIcapabilitiesareadvancingataremarkablepace—withsomearguingthatartificialgeneralintelligence(AGI)maybeachievedsoon.Yetwhenwelookattraditionaleconomicstatistics,weseeonlyupstreaminvestmentindatacenters,whiledownstreamimpactsfromthisrevolutionremainnearlyinvisible.GDPgrowthintheUnitedStatesandotheradvancedeconomieshasremainedmoderate,andproductivitystatisticshavebarelytickedup.Thequestion“whenwillweseeAIintheGDPstatistics?”hasbecomearecurringthemeineconomiccommentary.Onenaturalresponseispatience:AIadoptiontakestime,andtransformativeeconomicefectsmaysimplylieahead.Thisisalmostcertainlypartofthestory.
Butwebelievethereisanadditional,complementaryissueworthtakingseriously.Nationalaccountsweredesignedforaneconomyinwhichallproductionisultimatelyorganizedaroundhumansasthecentralpointofvaluecreation.Thiswasanentirelyappropriatedesignformostofeconomichistory,anditcontinuestoserveitscorepurposewell.However,therapidgrowthoftheAIsectorintroducesmeasurementchallengesthatexistingstatisticalcategorieswerenotbuilttoaddress.Thedi岱cultyisthatAIactivityishardtoseethroughthelensoftraditionalnationalaccountsandinthewayswetypicallymeasureGDP.
Thechallengeoperatesthroughseveralchannels.First,AIactivityisscattered:spendingonAIcompute,modeldevelopment,andAI-poweredservicesisspreadacrossdozensofindustrycategories—dataprocessing,cloudcomputing,softwarepublishing,professionalservices—makingitdi岱culttotracktheAIeconomyasacoherentwhole.Second,AIqualityimprovementsareunusuallyrapid:thepaceofimprovementinAIcapabilitiesisfarfasterthaninmostsectorsforwhichstatisticalagencieshavedevelopedqualityadjustmentmethods,raisingquestionsaboutwhetherstandardhedonictech-niquescapturewhatishappening.Third,AI’sroleintheeconomyisevolving:asAIsystemsbecomemorecapable,theymaytransitionfrombeingoneamongmanyinterme-diateinputstoplayingamorecentralroleinproduction,potentiallystrainingcategoriesthatweredesignedforaworldinwhichmachinesarepassivecapitalratherthanactivecontributorstoeconomicoutput.
AIisthelatestinaseriesofnewrapidly-growingtechnologiestointroducenewmeasurementchallenges,suchassemiconductorsandtheinternet.Butincontrastwithpreviousepisodes,AIwon’tnecessarilybeconstrainedbythesupplyofcomplementaryhumanlabor,whichcouldopenthedoortomuchlargermismeasurement.
Thesemeasurementchallengesmattertoday,andtheywillmattermuchmoreinthenearfuture.IfAIcapabilitiescontinuetoadvancerapidly,policymakersandresearcherswillneedtoolsthatcantracktheAIeconomy’sgrowthalongsidetraditionalstatistics.
4
Thecaseforbuildingsuchtoolsnow—whiletheAIsectorisstillrelativelysmall—restsonthesimpleobservationthatstatisticalinfrastructuretakestimetodevelop,andwaitinguntilmeasurementgapsbecomeacutemeansarrivingtoolate.
Thispapermakestworelatedbutseparablecontributions.First,weconstructproduction-sideestimatesofU.S.AIoutputcombininginferenceandtrainingactivities,quality-
adjustedusingAPIpricesatfixedperformancelevelsandestimatesofalgorithmicprogress.TheseestimatesrevealthatrealAIoutputhasgrownatratesvastlyexceedingnominalspending—afindingthatcanbeincorporatedintoconventionalGDPstatisticsasahedo-niccorrection,withoutrequiringanyreorganizationofexistingnationalaccounts.Second,weproposeaframeworkfor“AIGDP”thattreatstheAIsectorasacoherenteconomicentitytradingwiththehumaneconomy,anduseittoproduceanintegratedpictureoftheAIeconomy’ssizeandgrowth.
2ApproachandContributions
Inthispaper,weconstructamacroeconomicestimateoftotalAIproductionfortheUnitedStatesandacorrespondingquality-adjustedAIproductionindex.Ourmeasuresaggregatethroughseverallayers.First,webeginwithestimatesofrawcomputecapacity,anchoredinelectricityusageprojections(
Pateletal.
,
2024
)anddata-centercharacter-istics(
EpochAI
,
2026b
),andcross-checkedagainstchipsalesdata(
EpochAI
,
2026a
).Second,wemaprawcomputetonominalspendingusingacollecteddatasetofGPU-hourrentalprices.Third,weconstructqualityadjustments:forinference,usingtheevolutionofAPIprices
1
atfixedbenchmarkperformance(
Demireretal.
,
2025
);fortraining,usingthepaceofalgorithmicprogress(
Hoetal.
,
2024
)asmeasuredbythecomputerequiredtoreachagivenlevelofmodelperformance.
WeusetheseestimatestodevelopanascentbroaderframeworkforcalculatingAIGDP—theportionofeconomicvaluecreationmorecloselyassociatedwithAIcompu-tationratherthanhumancomputation.ThisframeworkprovidesapotentialconceptualfoundationfortrackingAI’scontributiontotheeconomyonitsowntermsratherthanasanincidentalbyproductofhuman-centeredaccountingcategories,oferinganovelandcomplementaryperspectiveontheforcesshapingoureconomy.
RelationshiptoTraditionalNationalAccounting
Itisusefultospelloutconcretelyhowourapproachrelatestostandardnationalac-countingpracticeandwhereitprovidesadditionalinformation.OurmeasurementsofAI
1API(ApplicationProgrammingInterface)pricesrefertotheper-unitfeeschargedbyAIproviders—suchasOpenAI,Anthropic,andGoogle—foraccesstolargelanguagemodels(LLMs)viasoftwareinterfaces.Pricesaretypicallyquotedpermilliontokensprocessed,whereatokenisroughlyawordfragment.
5
productionrepresentrefinementsthatfollowthespiritoftraditionalnationalaccountingrulesbutadaptthemtoprovideaclearerpictureofAI:
TrackingtheAIsectorasacoherentwhole.Nationalaccountsorganizeeconomicactivitybyinstitutionalunitandindustry.AI-relatedactivityisthereforedistributedacrossmanycategories,includingcloudcomputing,softwareandprofessionalservices.OurapproachinsteadtrackstheAIeconomyasacoherentwhole—allcomputeproduc-tion,modeldevelopment,andinferenceoutput—regardlessofwhichindustryclassifi-cationitfallsunder.Thisisanalogoustohowtradeeconomistssometimesconstructaccountsforthe“tradablesector”orhowenergyeconomiststracktheenergysectoracrossstandardindustryboundaries.Severalofthemeasuresweconstruct,particularlyourestimatesofnominalAIcomputespendingandrawcomputecapacity,couldserveasbuildingblocksforanAIsatelliteaccountwithintheexistingnationalaccountingframe-work,providingastructuredviewofAIactivitywithoutrequiringchangestoheadlineGDPmethodology.ThisisimportantbecauseAImaysoonbecomeoneoftheprimarydriversofvaluecreation,makingitnecessarytohaveanintegratedpictureofitsimpact.
Moregranularqualityadjustment.Standardstatisticalagenciesapplyhedonicpriceadjustmentsconservatively.Quality-adjustedpricedeclinesof20–30%peryearinfast-movingtechnologysectorslikesemiconductorscountasoutliers.Ourinferencedefia-tordeclinesapproximately94%peryear—roughlya16-foldincreaseinquality-adjustedoutputforagivendollarspent.Thisrefiectsthecompoundingoftwoforces:contin-uedchipe岱ciencyimprovementsandrapidalgorithmicprogress.Thischallengeisnotwithoutprecedent:
Hausman
(
1999
)showedthattheBLStelecommunicationsCPIwasbiasedupwardbyroughly2.3percentagepointsperyearsimplybyfailingtoaccountfortheintroductionofcellulartelephoneservice,anewgoodthato岱cialstatisticswereslowtoincorporate.Theextenttowhichthisaggressiveadjustmentisappropriatede-pendsonthedegreetowhichbenchmark-basedperformancegainstranslateintoeconomicvalue—aquestionweacknowledgeasanimportantcaveat.Butwebelievetheexerciseisinformativepreciselybecauseithighlightshowmuchthechoiceofdefiatormethodol-ogymattersforourunderstandingofAI’seconomictrajectory.IfAIcrossesathresholdsuchasAGI,behindwhichitbecomesbroadlyusefulacrosstheeconomy,ourestimatesmayshedimportantlightontheresultingmacroeconomicimplications.Inthatcase,thedefiatorforAIGDPmaynolongerdeclineatitscurrentpace,andtherapidincreasesinAIproductionmaytranslatemoredirectlyintotraditionalGDPgrowth.
Treatingmodeldevelopmentascapitalformation.OurframeworktreatsAItrainingasinvestmentin“modelcapital”—intangibleassetsthatimprovethequalityoffutureinferenceoutput.Thisextendsthelogicalreadyembeddedinnationalaccounts
6
sincethe2008SNArevision,whichcapitalizedR&Dasintellectualpropertyproducts.AsAImodeldevelopmentgrowsinscale,statisticalagenciesmayfinditusefultoidentifyitasadistinctassetcategorywithintheexistingIPPframework.
Ourproposalfordevelopinganascentframeworkfor“AIGDP”representsamorefundamentalreorientationofnationalaccountingstatisticsbutmaybecomeincreasinglyusefulastheroleofAIinourworldgrows:
ANascentFrameworkforAIGDP.Instandardnationalaccounting,mostofAIproductioncountsasanintermediategoodthatnetsoutinthecalculationoffinalGDP.ThisobscurestherapidchangesoccurringintheAIsector.WeproposeanewframeworkthatseparatesouthumanGDPandAIGDP,wherethelatterconsistsofalleconomicactivitythatisdrivenbyAI-basedcomputation.Inputssuchaselectricity,compute,andmaintenancethatAIsystemsrequiretooperateareclassifiedasimportstotheAIeconomy,whereasinferenceoutputsthataresoldtothehumaneconomyarecountedasexpensesandthusimports.Modeltrainingcountsasinvestmentinmodelcapital.Thisdepartureismoreconsequentialthantheothers:itrefiectsaview—developedfurtherinourAIGDPframeworkinSection
5
—thatasAIsystemsbecomemorecapable,theboundarybetween“capitalbeingusedup”and“anagentproducingoutput”maybe-comeincreasinglyblurred.WepresentthisasanalternativelenstocurrentpracticesofcompilingGDPthatwillbecomemoreandmoreusefulasAIcapabilitiesgrow.
OurframeworkofAIGDPisdesignedtoremaininformativeacrossarangeofsce-nariosforAI’sfutureeconomicrole.StandardaccountshandleAIadequatelyfortoday’spurposesbecauseAIisarelativelysmallintermediateinput.ButifAI’sroleexpandssubstantially—whethergraduallyorthroughamorerapidtransition—thentrackingtheAIeconomyonitsowntermsprovidesanearlysignalthatcomplementswhattraditionalstatisticscanofer.
Weemphasizethatourapproachisintendedtocomplement,notreplace,traditionalGDPmeasurement.Standardaccountswilllargelycontinuetoservetheircorepurposeoftrackingeconomicwelfarefromahumanperspective,althoughtheymayhavetobeadjustedonthemargins.Ourmeasuresprovideadditionalvisibilityintothisfast-movingsector.
DataLimitationsandUncertainty.OurestimatesrepresentaninitialattemptattakingamacroeconomicviewoftheAIeconomy,anddatalimitationsmeanthatstrongassumptionswererequired.Theallocationofcomputebetweentrainingandinferenceislargelyunobserved,forcingustoleanonanecdotalevidencethatthereisaroughlyequalsplit.OurvisibilityintoAIcompanies’grossmarginsislimited.Andtherelationshipbetweenbenchmarkperformance—whichanchorsourqualityadjustments—andactualeconomicvalueremainsuncertain.Thesegapspointtowardvaluablecollaborationsbe-
7
tweenstatisticalagencies,AIcompanies,andresearchers.AstheAIeconomygrows,sodoesthecaseforsystematicmeasurementinfrastructure.
3Methodology
ThecentralfocusofourmeasurementefortsisAIproduction,whichencompassesthecreationofinferenceoutputs(tokengenerationfortherestoftheeconomy)andtheformationofmodelcapital.Weprovideanoverviewofourmeasurementmethodologyanddatasourceshere.Thein-depthmethodologyisdescribedinthe
technicalappendix
.Below,wewillalsouseAIproductiontodevelopanascentconceptualframeworkformeasuringAIGDP.
3.1MeasuringAIProduction
Figure
1
showsaconceptualoutlineofAIproduction.ThecentraldriverofAIproductionisthegenerationofAIcompute,whichrequiresdatacentercapital(i.e.,GPUs)andtheelectricitytorunthem.Computeisthenallocatedbetweeninferenceandtraining.InferencecomputeisthecomputeusedtoproduceAIoutputsforuseinothertasks.InferencecomputeiscombinedwithAImodels(intangiblecapital)toproduceinferenceoutputs.TrainingandR&DcomputeincludesallcomputeusedintheAIR&Dprocessincludingpretrainingandpost-trainingofAImodels,experiments,andsyntheticdatageneration.TrainingcomputeiscombinedwithR&DalgorithmstoproduceAIR&D,whichleadstotheaccumulationofintangiblemodelcapital.Inthisway,trainingoutputimproveswiththeimprovementofR&Dalgorithms,andinferenceoutputsimprovewiththedevelopmentofnewmodelcapital.
8
Figure1:ConceptualoutlineofAIproduction.Computeisthecentralinput,allocatedbetweeninferenceandtrainingactivities,eachofwhichbenefitsfromqualityimprove-mentsovertime.
Power-basedestimatesofU.S.AIcompute.Ourprimarymethodologyforcal-culatingAIcomputeproductionfortheUnitedStatesstemsfromtheinsightthat,foragivencompositionofthestockofAIchips,computeoutputscaleswithpowerusage.Startingfromelectricityusageprojectionsfrom
Pateletal.
(
2024
),weapplychip-levelcharacteristicsfromtheEpochAIMLHardwaredataset(
EpochAI
,
2024
)—notablyThermalDesignPower(TDP)andbit-levelprocessingperformance—togetherwithGPUcapital-stocksharesfromtheEpochAIGPUClustersdataset(
EpochAI
,
2026b
),toconverttotalfacilityenergyintotheassociatedworkingtimeinGPU-hoursforeachchiptype.WethenapplyhourlyrentalratesforGPUs,collectedindependently,toobtainestimatesoftotalnominalspendingonAIcompute,onanimputedrentalpricebasis
2
.Thesamecharacteristicsletuscomputetotalphysicalcomputeproduction(inFLOPsorH100-equivalents)
3
.Derivationsandassumptionsbehindthesecalculationsareprovidedinthe
onlinetechnicalappendix
.
Asacross-check,foranalternativeestimateofcomputeproduction,weleveragere-centlyreleaseddatafromEpochAIonglobalchipsalesfromleadingproducers(
EpochAI
,
2026a
).Byassumingaconstantusageintensityandaccumulatingquarterlydeploymentsintoacumulativestock,weobtainabottom-upestimateofAIcomputespendingwith
2Assimplifyingassumptions,weholdrentalpricesconstantovertimeforagivenGPUtype,andweapplythelowest-availablecontractrateforeachgivenGPUtype,asmuchlarge-scalecomputeispurchasedthroughlong-termprivatecontracts.
3ItshouldalsobenotedthatFLOP-basedcomputemeasureslikelyunderstateimprovementsinchipe优ectivenessastheyfailtocaptureimprovedmemorycapacityandconnectionbandwidthinnewerchipgenerations.
9
globalscope,complementingtheU.S.-onlypower-basedmethod.Thetwoapproachesdrawonlargelyindependentdata,andtheiragreement—onceadjustedforthediferenceingeographiccoverage—providesaninformalvalidationofthepower-basedestimates.
Trainingvs.inferenceallocation.Dataonhowtoattributecomputebetweentrain-ingandinferenceisextremelylimited.Wethereforeimposetheassumption,basedonnarrativeaccounts,thatphysicalcomputeissplitroughlyequallybetweeninferenceserv-ingandR&Dactivities.MoredataonthissplitwouldbeveryvaluableforresearchersaimingtounderstandAIactivityatamacrolevel.
Inferenceoutputanditspricedefiator.RawinferencespendingdoesnotcapturetherapidqualitygainsinAIoutput.Tobuildaquality-adjustedinferencedefiator,weusedatagenerouslysharedby
Demireretal.
(
2025
),whichrecordstheminimumpromptprice(permilliontokens)forthecheapestavailablemodelwithineachintelligenceperformancetieronOpenRouter,observedweekly.WecomputeachainedFisherpriceindexacrossconsecutivemonthlypairs,includingonlytiersalreadypresentinthepriormonth—soanewlyintroducedfrontiertierenterstheindexthemonthafteritsintroduction,ratherthanartificiallyinfiatingit.Inthisway,wecapturepricemovementswithinperformancetiers,averagedacrossallavailableintelligencelevelsinagivenperiod.Thisyieldsatrenddeclineinper-tokenpricesofapproximately97%peryear(a35×e岱ciencygain),asillustratedinFigure
2
.Weadjustthisindexbytheapproximately2.2×annualtrendincreaseinbenchmarkresponselengths(
Embersonetal.
,
2025
),leadingtoanetinferencedefiatorthatdeclines94%peryear(roughlya16×fallintheefectivepriceofAIinferenceoutput).Asdocumentedby
Demireretal.
(
2025
),thisdeclineisbroad-basedacrossintelligencetiersandmodeltypes,consistentwithbroad-basedalgorithmicimprovementsovertimeallowingsmallermodelstomatchtheperformanceofolderlargermodels.
10
Figure2:ChainedFisherPriceIndexforAIinferencetokensatconstantcapability.Theindextrackstheminimumpricepermilliontokenswithineachintelligencetierandchainsthemintoasinglequality-adjustedseries.Therapiddeclinerefiectsfallinginferencecostsatfixedmodelcapability,notashifttowardlower-capabilitymodels.
Trainingoutputanditspricedefator.Toprice-adjusttrainingproduction,weapplythepaceofalgorithmicimprovementestimatedby
Hoetal.
(
2024
),whofindthatthecomputerequiredtotrainamodelatfixedperformancefallsbyroughlytwothirdseachyear—anannualizede岱ciencygainofapproximately65%,ora3×increaseinefectivecomputeperunitofphysicalcompute.
Onecaveatisthatweuseamixedmethodologytoaccommodatelimitationsintheavailabledata:weestimatenominalspendingontheproductionside,andcombineitwithapricedefiatorbasedonthepriceandcharacteristicsofinferenceoutputs.Thisassumesthatinferenceproviders’grossmarginsdidnotmateriallychangeovertime.
Sevillaetal.
(
2026
)findthatAIcompanieshavepositivebutmodestgrossmarginsoninference;sincethisisconsistentwithrelativelystablemargins,webelieveourestimatescapturetheappropriatefirst-orderdynamicsdeterminingtotalgrowthrates.
4Results
ThissectionpresentsourmainempiricalfindingsfortheUnitedStatesAIeconomyfrom2023to2025.
Table
1
reportsannualnominalcomputespendingandphysicalcomputeoutputfortheUSAIeconomy,basedonthepower-basedmethodologydescribedinSection
3.1
.
4
4Globalestimatesbasedonthechip-salesmethodologyshowmorerapidgrowth,withnominalspend-inggrowing291%in2024and166%in2025.Forfurtherdetails,seethe
onlineappendix
.
11
Nominalcomputespendinggrewatroughly144%peryearin2024and2025.Physicalcomputeoutput—measuredinH100-equivalentunits—grewevenfasteratroughly213%peryear,refiectingboththedeploymentofmorechipsandthetransitiontohigher-performancehardware.ContinuedimprovementsinAIchipe岱ciencyfromMoore’sLawmeantthatagivendollarcouldbuymoreFLOPs.Assuch,physicalcomputeproductionoutpacednominalspending.
4.1Quality-AdjustedAIProduction
RapidprogressonAIalgorithmsmeansthateachunitofcomputecouldbeusedtoproducedrasticallymoreAIoutput.Table
2
reportsquality-adjustedgrowthratesforinference,training,andaggregateAIoutput.ApplyingtheinferenceandtrainingdefiatorsdescribedinSection
3.1,
therealvolumeofAIinferenceoutputgrewroughly39×peryear.Quality-adjustedAIproductiongrowthcombinesinferenceandtraininggrowthweightedbytheirnominalshares,withgrowthinourfinalquality-adjustedAIProductionIndexsurpassing2000%peryear.
5FromAIProductiontoAIGDP
Inthissection,wegobeyondthetraditionalnationalaccountingconceptsanduseourmeasureofAIproductiontoconstructanovelframeworkforAIGDP,whichwedefine
12
asalleconomicvaluecreationassociatedwithAIcomputationratherthanhumanbrain-power.Toillustrate,imaginethatallAIactivities(datacentersandR&Dlabs)werelocatedonaseparateisland,withallhumanlaborandotherinputslikeelectricitybeingimportedfromoftheisland.IfthisAIislandwasitsowncountry,thenitisstraight-forwardtocalculatetheisland’sGDPbyaddinguptotalproductionoffinalproductontheisland(i.e.,excludingintermediateinputs)andsubtractinganyimportedinputsfromoftheisland,i.e.,fromthehumaneconomy.Thisquantity,whichwecall“AIGDP,”conceptuallycorrespondstothecontributionofAItotheGDPoftheoveralleconomy.
Figure3:EconomicfiowsintheAIeconomy.Importedinputs(left,blue)combinetoproducefinalproducts(right,red).AIGDPequalsfinalproductsminusimports.
Figure
3
outlinesthekeyeconomicfiowswithintheAIeconomy.AIProduction—theprocessofcombiningimportedelectricitywithAIchipstoproduceAIcomputewhichisthenappliedtoeitherinferenceorAIR&D—formsthecoreoftheeconomy.Alongthetopofthefigure,AIproductionisaugmentedbyincludingconsiderationofthemarginaccruingtocompaniessellinginference-basedAIservices.TheAIservicesmarginrepresentsthereturntointangiblecapitalembodiedinAImodelintellectualproperty(IP)ownership.Wealsonote“imported”laborinputswhichnetoutfromAIGDP.
Inourframework,mostinference-basedAIservicesare卩exported卩fromtheAIecon-
13
omy,withtheexceptionoftheuseofAItoolsintheAIR&Dprocessitself,whichremainsontheisland.AIR&Dandtrainingrepresentinvestmentinintellectualproperty(IPP)andareconsideredtheformationof卩modelcapital.卩ThisisconceptuallysimilartothetreatmentofR&DincurrentNationalAccounts
5
.ThesecontributionsdirectlyincreaseAIGDP,thoughcurrentmethodsarenotdirectlydesignedtocaptureAIefects,andtheintermediatenatureofAImakesitlessvisibleinexpenditureaccounts.
PhysicalcapitalformationconsistsoftheinstallationofAIchipsanddatacenters.AsseeninthebottomsectionofFigure
3
,thesearecurrentlycreatedinthehumaneconomyandare卩impo
温馨提示
- 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
- 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
- 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
- 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
- 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
- 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
- 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
最新文档
- 2026年区块链应用操作员考试题及答案
- 2026年广西柳州市初中学业水平考试模拟物理试题附答案
- 《运筹学》课件 第2章 单纯形法
- MySQL数据库技术与项目应用教程(微课版)(AI助学)(第3版)-习题答案 项目5
- 2026年湖南省醴陵市高二历史上册期末考试检测卷附答案【预热题】
- 2026年江苏省镇江市中考语文二模试卷
- 财务大数据分析电子教案
- 2026安阳六院面试题目及答案
- 数控钻工风险识别测试考核试卷含答案
- 香料合成工发展趋势测试考核试卷含答案
- 2026年广西继续教育公需科目试题及答案
- 2026年玉溪市中医医院公开招聘编外工作人员(17人)笔试备考试题及答案解析
- 政治+答案【一六八最后一卷】安徽合肥市第一六八中学等校2026届高三年级最后一卷(5.14-5.15)
- 山东省东营市2026年中考三模物理试题(含答案解析)
- 摩根士丹利 -半导体:中国AI加速器-谁有望胜出 China's AI Accelerators – Who's Poised to Win
- 市政设施损坏快速维修与抢修方案
- 2025-2026学年北师大版七年级数学下册期中达标测试卷(含答案)
- 灯火里的中国混声四部合唱谱冯
- 电动汽车充电桩建设合同能源管理协议
- AQ3062-2025《精细化工企业安全管理规范》专项检查表汇编(共5份)28
- 腐蚀检测技术
评论
0/150
提交评论