socialand.

policy··

Econ。mico●

rethunk

ExpertInsights

PERSPECTIVEONATIMELYPOLICYISSUE

VEGARDM.NYGAARD,ANUJINNERGUI,JONATHANW.WELBURN

Macroeconomic

Implicationsof

ArtificialIntelligence

August2025

PE-A3888-3

ThispublicationhascompletedRAND’sresearchquality-assurance

processbutwasnotprofessionallycopyedited.

Formoreinformationonthispublication,visit

/t/PEA3888-3

.

AboutRAND

RANDisaresearchorganizationthatdevelopssolutionstopublicpolicychallengestohelpmakecommunitiesthroughouttheworldsaferandmoresecure,healthierandmoreprosperous.RANDisnonprofit,nonpartisan,andcommittedtothepublicinterest.TolearnmoreaboutRAND,visit

.

ResearchIntegrity

Ourmissiontohelpimprovepolicyanddecisionmakingthroughresearchandanalysisisenabledthroughourcorevaluesofqualityandobjectivityandourunwaveringcommitmenttothehighestlevelofintegrityandethicalbehavior.Tohelpensureourresearchandanalysisarerigorous,objective,andnonpartisan,wesubjectourresearchpublicationstoarobustandexactingquality-assuranceprocess;avoidboththeappearanceandrealityoffinancialandotherconflictsofinterestthroughstafftraining,projectscreening,andapolicyofmandatorydisclosure;andpursuetransparencyinourresearchengagementsthroughourcommitmenttotheopenpublicationofourresearchfindingsandrecommendations,disclosureofthesourceoffundingofpublishedresearch,andpoliciestoensureintellectualindependence.Formoreinformation,visit

/about/research-integrity

.

RAND’spublicationsdonotnecessarilyreflecttheopinionsofitsresearchclientsandsponsors.

PublishedbytheRANDCorporation,SantaMonica,Calif.

?2025RANDCorporation

RANDeisaregisteredtrademark.

LimitedPrintandElectronicDistributionRights

Thispublicationandtrademark(s)containedhereinareprotectedbylaw.ThisrepresentationofRANDintellectualpropertyisprovidedfornoncommercialuseonly.Unauthorizedpostingofthispublicationonlineisprohibited;linkingdirectlytoitswebpageonisencouraged.PermissionisrequiredfromRANDtoreproduce,orreuseinanotherform,anyofitsresearchproductsforcommercialpurposes.Forinformationonreprintandreusepermissions,visit

/about/publishing/permissions

.

AboutThisPaper

iii

Astheworld,andtheUnitedStatesinparticular,confrontsthechallengesofeconomic

mobility,healthandlongevity,andnewfrontiersoftechnologyandhumanwell-being,havewedesignedoursocialandeconomicsystemstoeffectivelyaddresstheseconvergences?RANDisaddressingthesequestionsthroughaneffortcalledtheSocialandEconomicPolicyRethinktotransformtheapproachtosolvingsocialandeconomicpolicyproblems.Thebroadervolume

encompassessevenpublicationsthatdescribekeysocialandeconomicpolicystakesforpolicyonartificialintelligence(AI)adoption,examinehowpolicymakerscanaccountforrapidAI

capabilitiesdevelopmentandadoptioninandacrosssectors,andshowcasehowindustryand

policyleaderscanresponsiblymeetthechallengesofaneraoftransformativeAI.Thisreport

speakstotheboundsandconditionsforAIadoption,byexploringthepromiseofenhanced

productivitygainsfromAIthatcouldelevateincomesandthepotentialforsubstantiallabor

displacement,particularlyinsectorswithhighriskofautomationfromAI.Thispiecewillbeofinteresttopolicymakersatalllevels,aswellasindustryandcommunitystakeholders.

RANDSocialandEconomicWell-Being

RANDSocialandEconomicWell-BeingisadivisionofRANDthatseekstoactivelyimprovethehealthandsocialandeconomicwell-beingofpopulationsand

communitiesthroughouttheworld.

Funding

FundingforthisresearchwasprovidedbythecontributionsoftheRANDSocialand

EconomicPolicyAdvisoryBoard;generousgiftsbyFrankM.Clark,MichaelJ.Critelli,WilliamA.Downe,JiheeKimHuhandPeterYunHuh,andtheDonaldM.JamesFamilyFoundation;

andincomefromtheoperationofRANDSocialandEconomicWell-Being.

Acknowledgments

WethanktheentireSocialandEconomicPolicyRethinkteamfortheirinsightsthroughout

thisproject.Thepaper’speerreviewers,KeithCraneandSalilGunashekar,providedthoughtfulandinsightfulcommentsonanearlierdraft,whichledtosubstantialimprovements.Weextend

ourappreciationtotheleadershipofsocialandeconomicpolicyatRAND,AnitaChandra,PeterHusseyandHeatherSchwartz;totheRANDSocialandEconomicPolicyboardfortheirsupportandadvice;tothemembersoftheRethinkfinancialservices,healthcare,andclimateandenergy

iv

workinggroupsfortheiradviceandscoping;toLisaCoeforhercarefuledits;andtothemembersoftheRethinkteamwhoseinputhelpedguidethisanalysis.

Summary

v

TheincreasingintegrationofArtificialIntelligence(AI)intotheeconomypresentsbothchallengesandopportunities,withsignificantimplicationsforeconomicgrowthand

employment.Inthisperspective,weexplorethedualeffectsofAI:thepromiseofenhancedproductivitygainsthatcouldelevateincomes,andthepotentialforsubstantiallabor

displacement,particularlyinsectorswithhighriskofautomationfromAI.1

While“AI”isoftenreferencedinthesingular,thetermencompassesabroadandgrowing

rangeofcapabilitiesaswellasthewaysuchcapabilitiesareutilizedindifferentcontexts.We

focusontheadoptionofthebroadclassofAItechnologiesinthisperspectivethatencompassesMachineLearning,DeepLearning,NaturalLanguageProcessing,LargeLanguageModels,andGenerativeAI.WhileAIwillimpactallcountries,thisperspectivefocusesexclusivelyonthe

impactontheU.S.economyandtheU.S.labormarket.Thus,wedonotexamineglobalimpactsofAI,includinghowAIadoptionvariesacrosscountriesortheimplicationsofAIoncross-

bordercapitalandlaborflows.

WefirstdiscusstheuncertainimpactofAIonfutureproductivitygrowthandthenusean

economicgrowthmodeltoexaminetheimpactofdifferentproductivitygrowthscenariosfortheU.S.economy,includingimplicationsforgrossdomesticproduct(GDP)percapitaandfederal

governmentdebt.Next,weutilizedatafromtheBusinessTrendsandOutlookSurvey(BTOS)toassessthecurrentlandscapeoftheadoptionofAIacrossvarioussectors.Usingrecentresearch

findings,weidentifyoccupationalgroupsthataremostsusceptibletoautomationstemmingfromAI.Despiteconcernsoverjoblosses,preliminaryfindingssuggestthatAImayinitially

complementratherthanreplacehumanlabor,althoughthistrendcouldshiftasthetechnologyadvances.

IntheeventthatAIleadstosignificantlabordisplacement,policymakersmaywantto

considerusingactivelabormarketprograms,suchasworkforceretrainingprograms,tomitigatelong-termincomelossesfordisplacedworkers.Wereviewtheliteratureonworkforceretrainingprograms,whichfindsmixedevidenceontheirefficacy.Forexample,trainingprogramstendtobelesseffectiveforolderworkers,andreturnsareoftenhigherforquantitativecourses.These

findingssuggestthatAI-inducedlayoffsmaytriggersignificantincomelossesforolderworkers,withlong-termnegativeimplicationsforwealthaccumulationandfinancialpreparednessfor

retirement.

1Fordetails,seeMauroCazzanigaetal.,"Gen-AI:ArtificialIntelligenceandtheFutureofWork,"IMF,2024;andFrancescoFilippuccietal.,"TheimpactofArtificialIntelligenceonproductivity,distributionandgrowth:Key

mechanisms,initialevidenceandpolicychallenges,"OECDArtificialIntelligencePapers,2024.

vi

OurfindingsunderscoretheimportanceofproactivemeasurestomonitorAI’simpactonproductivity,labormarkets,andtheoverallhealthoftheU.S.economy.Weadvocatefor

improveddatacollectiontoinformfuturestrategiestoaddresseconomicissuesrelatedtoAI.

Additionally,policymakersmaywanttoexaminewhichretrainingprogramsaresuccessfulat

increasingthelikelihoodofreemploymentfollowinganAI-inducedlayoff,includingthespecificcoursesthataregiven,andshouldalsoexaminehowtheefficacyvarieswithworker

characteristicssuchasage,educationalattainment,occupation,andsectorofemployment.

KeyFindings

Thekeyfindingsfromouranalysisandreviewofrecentliteratureareasfollows:

?AImayinducehigherproductivitygrowth:WhilerecentresearchsuggestthatAImayacceleratefutureproductivitygrowth,thereisnoconsensusonthemagnitudeoftheimpact.TherangeofimpactspredictedbydifferentstudiessuggestthatAImightadd0.1–1.5percentagepointstoannualproductivitygrowthintheU.S.overthe

nextdecade.

?AI-inducedhigherproductivitygrowthmayincreaseper-capitaGDP:Realper-capitaGDPcouldbenearly$7,000higherby2035intheeventthatAIsuccessfullyincreasesfutureproductivityby0.5percentagepointsperyear.

?Datacenterconstructionmaysupportmillionsofjobs:Theconstruction,

maintenance,andoperationofdatacentersmaygeneratesubstantialdirectand

indirectregionaleconomicbenefits,supportmillionsofjobs,andcontributebillionstoannualGDP.However,theprojectedcomputationalresourcesrequiredtooperatethedatacentersandtrainfutureAIsystemsmightrequiresignificantamountsof

powergenerationthatcouldcomplicatethegoalofdecarbonizingenergyproduction.

?EarlyadoptionofAIappearstohavecomplementedworkerproductivity,butfuturelabormarketimpactsareuncertain:AsofFebruary2024,surveydata

indicatethatAIwasmorelikelytoincreaseratherthandecreaseemployment,

suggestingthatAIwascomplementingworkerproductivityratherthansubstitutingforhumantasksacrossawiderangeofsectors.However,impactsmaychangeasAIcapabilitiescontinuetoadvanceandadoptioncontinuestodeepen.

?MoredatacollectionisnecessarytounderstandtheimpactofAIonlabor

marketsandtheefficacyofworkforceretrainingprograms:PolicymakersmaywanttomonitortheeffectsofAIonlabormarkets,includingtrackingwhetherAIisimprovinglaborproductivity,examininghowAIisbeingemployedindifferent

sectors,trackingwhattypeofoccupationsarebeingdisplacedbyAI,andexamininghowlongittakesworkersthathaverecentlybeendisplacedbyAItofindanewjob.AsAIpresentsnewlabormarketchallenges,moreworkwillalsobeneededto

understandwhetheractivelabormarketpolicies,includingworkforceretraining

programs,areeffectiveatincreasingthereemploymentlikelihoodforworkersthatweredisplacedbyAI.

Contents

vii

AboutThisPaper iii

SocialandEconomicWell-Being iii

Funding iii

Acknowledgments iii

Summary v

KeyFindings vi

Figures viii

MacroeconomicImplicationsofArtificialIntelligence 1

DualEffectsofArtificialIntelligence 1

Takeaways 21

Abbreviations 22

References 23

Figures

viii

Figure1.ProjectedAI-DrivenGrowthinRealGDPperCapita(constant2024dollars) 4

Figure2.ProjectedAI-DrivenReductionintheFederalDebtBurden 5

Figure3.DataCenterConstructionSpending(%ofGDP) 6

Figure4.NumberofDataCentersbyState 7

Figure5.PercentChangeintheShareofDataCenterEmploymentinTotalEmployment

(2016Q2–2024Q2) 8

Figure6.ShareofBusinessesbySectorCurrentlyUsingAItoProduceGoodsandServices 12

Figure7.ShareofBusinessesbySectorReportingChangesinEmploymentDuetoAIAdoption

15

Figure8.ShareofBusinessesbySectorwhereAIhasReplacedTasksPreviouslyPerformedby

Workers 16

Figure9.ShareofBusinessesbySectorPlanningtoReplaceHumanTaskswithAI 17

Figure10.PercentagePointIncreaseintheShareofBusinessesbySectorPlanningtoReplace

HumanTasksbyAI 18

MacroeconomicImplicationsofArtificialIntelligence

1

DualEffectsofArtificialIntelligence

TheincreasingintegrationofArtificialIntelligence(AI)intotheeconomypresentsboth

challengesandopportunities,withsignificantimplicationsforeconomicgrowthand

employment.Inthisperspective,weexplorethedualeffectsofAI:1)thepromiseoflarge

productivitygainsthatcouldraiseaverageincomes;and2)thepotentialforAI-induced

automationofworkertasksthatcouldbringaboutsignificantjoblossesacrossmultiple

economicsectors,includingtheservicessector,whichcurrentlymakesupnearly80percentofthetotalU.S.workforce.

While“AI”isoftenreferencedinthesingular,thetermencompassesabroadandgrowing

rangeofcapabilitiesaswellasthewaysuchcapabilitiesareutilizedindifferentcontexts.Inthisperspective,wefocusontheadoptionofthebroadclassofAItechnologiesthatencompasses

MachineLearning(ML),DeepLearning(DL),NaturalLanguageProcessing(NLP),LargeLanguageModels(LLMs),andGenerativeAI(GenAI).2

ThePromiseofAI-InducedProductivityGrowth

Growthintotalfactorproductivity(TFP)—theaveragereal(inflation-adjusted)outputper

unitofcombinedlaborandcapitalservices—haslongbeenamajorengineofeconomicgrowthintheU.S.—alongwithgrowthinboththecapitalstockandthelaborforce3—andhasbeena

keycontributortoincreasesinper-capitaincomelevels.However,TFPgrowthhasbeenona

downwardtrajectoryinrecentdecades,aslowdownthatisprojectedtocontinuegoingforward.TheCongressionalBudgetOfficeforecasts1.1percentannualgrowthinTFPduringthenext

threedecades,downfromanaverageof1.3percentduringtheprecedingthreedecades.4Severalreasonscontributetothisprojectedproductivityslowdown,includingaslowdowninthegrowthofeducationalattainmentandlowerfederalgovernmentspendingasashareofGDPon

2Fordetails,seeJonathanW.Welburn,MayaBuenaventura,VegardM.Nygaard,ChandraGarber,LeahDion,

PedroNascimentodeLima,AntonShenk,AnujinNergui,BenjaminBoudreaux,CarterPrice,BebaCebralic,SeanMann,FlanneryDolan,andKarishmaV.Patel,RethinkingSocialandEconomicPolicyintheAgeofGeneral-

PurposeArtificialIntelligence:NavigatingtheTradeoffsofAIAdoption,RANDCorporation,RR-A3888-2,forthcoming.

3TFPhasaccountedfor34percentofaveragegrowthinoutputperworkeracrossWesternCountries(ScottL.Baier,GeraldP.DwyerJr.,andRobertTamura,“HowImportantAreCapitalandTotalFactorProductivityforEconomicGrowth,”EconomicInquiryVol.44:1,2006).

4CBO,"TheLong-TermBudgetOutlook:2025to2055,"2025b.

2

infrastructure,education,training,andR&D.5Thisprojectedproductivitygrowthslowdowncouldhavesignificantnegativeeconomicandfiscalimplications,includingaslowdowninthegrowthofper-capitaGDP—themostcommonlyusedproxyforaveragelivingstandards6—

slowergrowthintaxrevenues,andincreasedfederalindebtednessiffiscalpoliciesarenotadjustedtocompensateforlowerrevenues.7

ImpactofAI-InducedHigherProductivityGrowthonper-capitaGDPandFederalDebt

WhileongoingadvancesinAIhasthepotentialtoacceleratefutureTFPgrowth,thereisnoconsensusonthemagnitudeofthatimpact.SomestudiessuggestthatAIwillleadtoasharp

accelerationinproductivitygrowth,potentiallyboostingU.S.laborproductivityby1.5

percentagepointsperyear,8althoughbarrierstoadoptionmaydelaygrowth.9Otherreports

predictthatAIwillhavemore-modestimplicationsforfutureproductivitygrowth.Basedona

task-basedaggregationframework,AcemoglupredictsthatAIwilladdlessthan0.1percentagepointstoannualU.S.TFPgrowthoverthenexttenyears.10AghionandBunelobtainmore

optimisticpredictions,estimatingthatAIwillincreaseaggregateproductivitygrowthby0.8–1.3percentagepointsperyearoverthattimeperiod,andFillippuccietal.predictthatAIwilladd

0.4–0.9percentagepointstoannuallaborproductivitygrowth.11

GiventheimportanceoftheprojectedgrowthinTFPforthefutureeconomicsituationintheU.S.andthegovernment’sfiscaloutlook,weexaminetheimplicationsofacceleratingfuture

TFPgrowthbeyondthe1.1percentannuallong-rungrowthrateassumedunderCBO’sbaselineprojection.Specifically,weconsideranAIscenarioinwhichU.S.TFPgrowsbyanadditional0.5percentagepointsperyear.WhiletheAIscenariorepresentsasharpincreaseinTFPgrowth

5SeeVollrathandCBOfordetailsonthereasonsfortheprojectedslowdowninTFPgrowth(DietrichVollrath,

"FullyGrown:WhyaStagnantEconomyIsaSignofSuccess,"UniversityofChicagoPress,2020.;andCBO,"TheBudgetandEconomicOutlook:2025to2035,"2025a).

6SeeJonesandKlenowaswellasFalcettoniandNygaardfordetailsonthestrengthsandweaknessesofGDPpercapitaasameasureofaveragelivingstandards(CharlesI.JonesandPeterK.Klenow,“BeyondGDP?Welfare

AcrossCountriesandTime,”AmericanEconomicReview,Vol.106:9,2016;andElenaFalcettoniandVegardM.Nygaard,“AComparisonofLivingStandardsAcrosstheUnitedStatesofAmerica,”InternationalEconomic

Review,Vol.64:2,2023).

7CBO,“HowChangesinEconomicConditionsMightAffecttheFederalBudget:2025to2035,”2025d.

8J.BriggsandD.Kodani,“ThePotentiallyLargeEffectsofArtificialIntelligenceonEconomicGrowth,”GoldmanSachsEconomicsResearch,2023.

9GoldmanSachs,“UpgradingOurLonger-RunGlobalGrowthForecaststoReflecttheImpactofGenerativeAI,”2023b.McKinsey&CompanyestimatethatworkautomationresultingfromthecombinationofGenAIandothertechnologiescouldadd0.5–3.4percentagepointsannuallytoglobalproductivitygrowth(MichaelChuietal.,“TheeconomicpotentialofgenerativeAI:Thenextproductivityfrontier,”McKinsey&Company,2023).

10DaronAcemoglu,“TheSimpleMacroeconomicsofAI,”EconomicPolicyVol.39(120),2024.

11PhilippeAghionandSimonBunel,“AIandGrowth:WhereDoWeStand,”FederalReserveBankofSan

Francisco,2024;andFrancescoFilippucci,PeterGal,andMatthaisSchief,“MiracleorMyth?Assessingthe

macroeconomicproductivitygainsfromArtificialIntelligence,”O(jiān)ECDArtificialIntelligencePapers,Vol.29,2024.

3

relativetothebaselineprojection,itiswellwithintheboundspredictedbytheliteratureonthepotentialimpactsofAIonfutureproductivitygrowth.Wehencedeemthis“highgrowth”AIscenarioasafeasible,albeitoptimistic,futurescenario,atleastintheshortrunasbarriersto

adoptionmightdelaytheaccelerationinproductivitygrowth.

WeuseageneralequilibriumeconomicgrowthmodeldevelopedbyCBOtoforecastthe

impactofhigherTFPgrowthonfutureGDPpercapita,federalgovernmentdeficits,andthe

interestrateonfederaldebt.12Thesevariablesenableustoforecastthefuturefederaldebt

burden,calculatedastheratiooffederaldebtheldbythepublictoGDP.Thisapproachhas

certainlimitations.Forexample,itassumesthatTFPgrowthincreasesduetoAI,butitdoesnotaccountforanyinvestmentsorpolicychangesthatmightbenecessarytoachievethisTFP

growth.Additionally,itdoesnotaccountforpotentialadverseimpactsonlabormarkets,asdiscussedmoreinthefollowingsection.Thus,thisanalysisonlyprovidesanestimateofthedirectimpactofAI-inducedproductivitygrowthontheeconomy.

Figure1plotstheprojectionsforrealpercapitaGDPmeasuredinconstant,2024dollars,

wherethesolidblacklinecorrespondstothebaselineforecastbyCBOandthedashedbluelinecorrespondstotheAIscenario(i.e.,higherTFPgrowth).AlthoughtheAIscenarioaddsonly0.5percentagepointstoannualTFPgrowth,theimplicationforrealper-capitaGDPisconsiderablebecauseofcompounding,wherebygainsinincomewillaccrue,withinterest,overtime.Thisisevidentbythedivergentpathsofthetwoprojections,whichshowsthatrealper-capitaGDP

couldbenearly$7,000higherby2035undertheAIscenariocomparedtothebaselinescenario.Notably,therealper-capitaGDPgapbetweenthesetwoscenarioswouldcontinuetogrowifAIleadstoasustainedaccelerationinTFPgrowthbeyondthe2035projectionwindowconsideredhere.

12CBO,"WorkbookforHowChangesinEconomicConditionsMightAffecttheFederalBudget:2025to2035,"2025c.

4

Figure1.ProjectedAI-DrivenGrowthinRealGDPperCapita(constant2024dollars)

SOURCE:RANDAnalysisusingCBO(2025b)andCBO(2025c).

NOTES:Thebaselineforecast—solidblackline—usesCBO’sbaselinepotentialTFPgrowthrateprojection.TheAIScenarioproducedbyRAND—dashedblueline—assumesthatpotentialTFPwillgrowbyanadditional0.5

percentagepointsperyearoverthistimeperiodbeyondtheTFPgrowthassumedunderCBO’sbaseline.WedeemthehighgrowthAIscenarioasafeasible,albeitoptimistic,futurescenariobasedontheboundspredictedbythe

literatureonthepotentialimpactsofAIonfutureproductivitygrowth.WeuseageneralequilibriumeconomicgrowthmodeldevelopedbyCBOtoquantifytheimpactofhigherTFPgrowthonrealGDPpercapita.Notably,theforecastsaresubjecttouncertaintybecauseofthelong,ten-year,projectionwindow.

Figure1showsthatAIhasthepotentialtoincreaseaverageincomesintheU.S.However,asthegainsfromAIarelikelytobeunevenlydistributedacrossindividuals,AIcould

simultaneouslyleadtobothhigheraverageincomesandhigherincomeinequality.Forexample,AcemoglupredictsthatAIwillwidenthegapbetweencapitalandlaborincome.13Asdiscussedmoreinthefollowingsection,AImayalsoleadtoreductionsinincomeforalargeshareof

individualsifAIresultsinsignificantlabordisplacement.

WenextexaminetheimpactofAI-inducedTFPgrowthonfuturefederaltaxrevenuesandexpenditures,alongwiththeimplicationsforfederalgovernmentindebtedness,assumingno

changesincurrentfiscalpolicies.Federaldebtheldbythepublicwasat98percentofGDPin2024.DatafromCBOshowthatthisisthesecond-highestleveloffederalindebtednessin

relationtoGDPthattheU.S.economyhaseverwitnessed,surpassedonlybytheexperience

13Acemoglu,2024.

5

followingWorldWarIIwhenfederaldebtclimbedto106percentofGDP.14AsillustratedbytheblacksolidlineinFigure2,federaldebt-to-GDPisprojectedtosurpassthatlevelinthenearfuture,withCBO’smostrecentforecastprojectingthatitwillreach118percentby2035.15

Figure2.ProjectedAI-DrivenReductionintheFederalDebtBurden

FederalDebtHeldbythePublic(%of

120

Projection

110

100

GDP)

90

80

70

2015

2016

2017

2018

2019

2020

2021

2022

2023

2024

2025

2026

2027

2028

2029

2030

2031

2032

2033

2034

2035

Baseline(CBO)AIScenario

SOURCE:RANDAnalysisusing(CBO,2025b)andCBO(2025c).

NOTES:Thebaselineforecast—solidblackline—usesCBO’sbaselinepotentialTFPgrowthrateprojection.TheAIScenarioproducedbyRAND—dashedblueline—assumesthatpotentialTFPwillgrowbyanadditional0.5

percentagepointsperyearoverthistimeperiodbeyondtheTFPgrowthassumedunderCBO’sbaseline.WedeemthehighgrowthAIscenarioasafeasible,albeitoptimistic,futurescenariobasedontheboundspredictedbythe

literatureonthepotentialimpactsofAIonfutureproductivitygrowth.WeuseageneralequilibriumeconomicgrowthmodeldevelopedbyCBOtoquantifytheimpactofhigherTFPgrowthonthefuturedebt-to-GDPratio.Notably,theforecastsaresubjecttouncertaintybecauseofthelong,ten-year,projectionwindow.

Weillustratetheimpactofthe“highgrowth”AIscenariobythedashedbluelineinFigure2.ThissustainedhigherTFPgrowthcouldleadtosubstantiallyhighertaxreceiptsanda

significantlylowerriseinthedebt-to-GDPratio:107percentofGDPby2035,comparedto118percentunderCBO’sbaselineprojection.Thiscouldhavelargepositivefiscalimplicationsfor

14FordetailsontheevolutionoffederaldebtsinceWorldWarII,alongwithestimatesonwhatwouldberequiredtoreducethefuturefederaldebtburden,seeVegardM.Nygaard,CarterPrice,andAkshayaSuresh,"StrategiesforReducingtheBurdenofFederalDebt:HowtheU.S.ReducedtheDebtBurdenFollowingWWII,AndWhatIt

WouldTaketoDoSoAgain,"RANDCorporation,WR-A4183-1,forthcoming.

15CBO,2025b.

6

theU.S.economy,includinga$2trillionnominalcumulativereductioninfederaldeficitsoverthecourseoftheten-yearprojectionwindowrelativetobaselineforecasts.16

AIInfrastructureInvestment

WithoutsubstantialimprovementsintheenergyefficiencyofAImodels,maintainingthe

projectedpaceofAIinnovationwillrequirelargeinvestmentsindatacenterstoaccommodate

thegrowthincomputingdemandrequiredtotrainandrunAImodels.AsshowninFigure3,datacenterconstructionasashareofGDPhasalreadyincreasedsubstantiallysince2021—oneyearbeforethereleaseofChatGPT—morethandoublingfromabout0.05percentin2021toabout

0.11percentin2024.Spendingondatacenterconstructionisprojectedtocontinuetoriseinthefuture,reachingroughly0.15percentofGDPbytheendof2026.

Figure3.DataCenterConstructionSpending(%ofGDP)

SOURCE:RANDAnalysisusinghistoricalandprojecteddataondatacenterconstructionspendingarefrom,respectively,theU.S.CensusBureau(2025c)andAIA(2025).HistoricalandprojecteddataonGDParefrom,respectively,U.S.BureauofEconomicAnalysis(NationalIncomeandProductAccounts,Table1.1.5)andCBO(2025a).

NOTE:DatacenterconstructionspendingprojectionsarebasedontheAIA(2025)ConsensusConstructionForecast,indicatinga21.9percentincreasein2025anda14.6percentincreasein2026.

16CBO,2025c.

7

ThisgrowthinAIinfrastructurehasbeengeographicallyconcentrated,withVirginia,Texas,andCaliforniahostingthelargestnumberofdatacenters(Figure4).Thesethreestatesalone

accountforaboutone-thirdofthe3,664datacenters,asofApril2025.

Figure4.NumberofDataCentersbyState

SOURCE:RANDAnalysisusingDataCenterMap(2025)

NOTES:Thisfiguredisplaysstateswith50ormoredatacenters.Thetotalcountofdatacentersacrossall50statesandtheDistrictofColumbiais3,664,asofApril2025.

Theconstruction,maintenance,andoperationofdatacenterscangeneratesubstantialdirectandindirectregionaleconomicbenefits.Forexample,Figure5showsthepercentagechangeintheshareofdatacenteremploymentintotalemploymentbetween2016Q2and2024Q2,with

statessuchas