Estimating Aggregate Data Center Investment with Project-level Data

Finance and Economics Discussion Series Federal Reserve Board, Washington, D.C. ISSN 1936-2854 (Print) ISSN 2767-3898 (Online) Estimating Aggregate Data Center Investment with Project-level Data Eirik Eylands Brandsaas, Daniel Garcia, Robert Kurtzman, Joseph Nichols, and Adelia Zytek 2025-109 Please cite this paper as: Brandsaas, Eirik Eylands, Daniel Garcia, Robert Kurtzman, Joseph Nichols, and Adelia Zytek (2025). “Estimating Aggregate Data Center Investment with Project-level Data,” FinanceandEconomicsDiscussionSeries2025-109. Washington: BoardofGovernorsofthe Federal Reserve System, https://doi.org/10.17016/FEDS.2025.109. NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Estimating Aggregate Data Center Investment with Project-level Data * Eirik Eylands Brandsaas†1, Daniel Garcia‡1, Robert Kurtzman§1, Joseph Nichols¶1, and Adelia Zytek||1 1Federal Reserve Board of Governors December 17, 2025 Abstract DatacenterinvestmentintheU.S.hasincreasedrapidlyinthepost-pandemicera,andplansfor futureinvestmenthavesurgedfurther. Forecastinginvestmentatsuchaturningpointisanimportant but potentially fraught exercise, especially given lags in aggregate data availability. We develop a straightforward method to forecast aggregate investment using project-level microdata and a small numberofparameters: specifically,abandonmentrates,timefromplan-to-start,andtimefromstart-tocompletion. Asakeyvalidationofourapproach,wegenerateestimatesthatmatchtherecenthistory of aggregate data center investment in the NIPAs. We then use our method to generate nowcasts of aggregatedatacenterinvestmentintheshortrun,withthemeanforecastindicatingthatinvestmentwill increaseto$370billionannualizedby2026:Q2. Wecanextendourmethodologyfurtherout,butour forecaststhenbecomeconditionalontheassumedflowofnewdatacenterplans. Assumingfutureplans rangefromone-fourthtotwicetheaveragepaceofplansfrom2024-2025impliesarangeofinvestment forecastsof$360billionto$930billionin2027,demonstratingthesubstantialupsideanddownside riskstofuturelevelsofinvestment. Keywords: datacenters,commercialrealestate,forecasting,construction,time-to-plan JELClassification: R33,E22,E32,L74 *WethankChrisKurzandRavenMolloyfortheirhelpfulfeedback. Theviewsexpressedinthispaperaresolely thoseoftheauthorsanddonotnecessarilyreflecttheopinionsoftheFederalReserveBoardortheFederalReserve System. †DivisionofResearch&Statistics,FederalReserveBoard;eirik.e.brandsaas@frb.gov. ‡DivisionofResearch&Statistics,FederalReserveBoard;daniel.i.garcia@frb.gov. §DivisionofResearch&Statistics,FederalReserveBoard;robert.j.kurtzman@frb.gov. ¶DivisionofResearch&Statistics,FederalReserveBoard;joseph.b.nichols@frb.gov. ||DivisionofResearch&Statistics,FederalReserveBoard;adi.k.zytek@frb.gov. 1

1. INTRODUCTION The post-pandemicera has seen asurgein data centerinvestment, with investment indata center structureshavingroughlyquadrupledfrom2021to2025. Meanwhile,announcedplansforfuture datacentershaveincreasedsignificantly,totalingover$1trillioninthefallof2025. Giventhese developments, an important question facing policymakers and practitioners is to understand the scaleanddynamicsofsuchinvestmentintheshortrunandoverlongerhorizons. Ourpaperprovidesamethodologyforansweringthisquestionthatcangenerateaggregatereal-time forecasts using microdata on project plans and a parsimonious number of empirically generated parameters. This approach provides an alternative framework for generating forecasts of nonstationary time series at the onset of an inflection point.1 We apply our method to project-level datafromthedatacentersector,whichisinthemidstofexactlysuchaninflectionpointandisof generalinterestgivenitsroleinthedevelopmentofpotentiallytransformativeAI-poweredservices. Our methodproduces reasonable out-of-sample performance in the shortrun, and we showhow it canbeextendedtolongerhorizons,highlightingtheimportanceofassumptionsontheflowofnew plansinthefuture. Togenerateanowcastoffutureinvestmentovertheshortrun,definedasone-yearout(thetypical timefromplan-to-startinourdataset),werequirethreeparameters. Forprojectsunderconstruction, werequirethetime-to-completion. Forprojectsinplanning,werequirethetimefromplan-to-start. Finally,weneedtheabandonmentprobabilityfortheproject,whichisabout33percenthistorically in the data center sector in our sample. To generate a forecast beyond one year, which we label as the medium run, one needs to make assumptions on the inflows of new plans. This is a more tenuoustask,asitrequiresassumptionsaboutthescopeandvalueoffutureplans.2 We apply our methodology to estimate and nowcast data center investment over both the recent historyandtheshortrunusingmicrodataonprojectplansfromtheDodgeConstructionNetwork 1Ourapproachprovidesaforecastofthemagnitudeandtimingoftheinflectionpointbeforeitisevidentinthe officialstatisticsbyusingannouncedplans,insteadofwaitinguntilthereareenoughdataaftertheinflectionpointto estimateamoretraditionalnon-stationarytimeseriesmodel,suchasregimeswitchingortime-varyingparameters models(StockandWatson,1996). 2To be precise as to how we are using our methodology over different time periods, we generate estimates of investmentfromthemicrodataovertheperiodwherewealsohavetheofficialaggregatedata. Wegeneratenowcasts ofinvestmentovertheperiodwhereweuseonlythecurrentlyreportedprojectplansbutdonotyethavetheofficial aggregatedata. And,lastly,wegenerateforecastsofinvestmentoverlongerperiodsthatrequireassumptionsoffuture planactivity. 2

(Dodge)andthenextendtheanalysisbygeneratingforecastsoverthemedium-term. Anadvantageof theDodgedataisthatitisusedasaninputintheCensus’ConstructionPutinPlace(CPIP)survey, which informs BEA’s investment measurements. Additionally, the project microdata typically includestotalcostestimates,whichallowsustospeaktooveralldatacenterinvestment,asopposed tosolelyinvestmentinthedatacenterstructure. WefirstusethehistoryofdatacenterprojectplansreportedinDodgefrom2003to2025togenerate estimatesofthethreekeyparametersdiscussedabove. ApplyingthesetothehistoryofDodgedata center projects provides a quantitatively similar estimate to aggregate data center investment in theNational Incomeand Product Accounts(NIPAs),providingavalidationofour approach. We thenusetheDodgemicrodata asofSeptember2025tonowcast thenear-termpathofdatacenter investment via a Monte Carlo simulation approach. In the mean of our simulations, investment increases from approximately $60 billion in 2024:Q4 to $180 billion in 2025:Q4, reaching an annual rate of roughly $370 billion by 2026:Q2. We then forecast investment through 2027 for three different scenarios of the future flow of new data center plans. These scenarios assume futureplans rangefrom one-fourthtodouble theaveragepaceof plansfrom 2024-2025. In these simulations, mean annualized investment in 2027 ranges from about $360 billion in the more pessimisticalternativeto$930billioninathemoreoptimisticscenario;effectively,bytheendof 2027,thelevelofinvestmentislargelydeterminedbytheassumedinflowofnewplansandnotthe currentstockofplans. Aspartofourwork,wedevelopanapproachforidentifyingdatacenterspecificinvestmentinthe NIPAs. Though aggregate data center investment in structures is already produced in the NIPAs, computerequipmentandperipheralsinvestment(E&I)isnotbrokenoutintodatacenterinvestment. Because such E&I investment includes other product categories, such as business purchases of laptops and other computer equipment, we argue that one should remove the level of investment not attributed to data centers. We present a simple method to do this to construct a measure of aggregatedatacenterinvestmentintheNIPAs. BecausetheNIPAscomeoutwithalag,anadditional advantageofour bottom-upapproachis thatitallowsone toestimatesuch aggregateinvestment in real-time. OurworkprovidesanimportantcomplementtoresearchonthebroadereconomicimpactofAI(see, e.g., Acemoglu 2025, Furman and Seamans 2019, or Brynjolfsson et al. 2025). While these papers are focused on the potential for AI to re-order the economy, primarily through increasing labor productivity,suchpotentialdevelopmentsareafunctionofthebuild-outoftheunderlyingphysical 3

infrastructureofdatacenters. Ourworkthusprovidesavaluablecomplementtothisliteraturewith an innovative method that can translate detailed project level planning data into a time series of actualinvestmentthatisapre-conditionfortheprovisionofnext-generationAIservices. This analysis also contributes to the literature studying investment using project-level planning microdata. Glancyetal.(2025)showthatabandonmentsoutofplanningarecrucialdeterminants ofshort-runsupplyelasticitiesinthecommercialconstructionspace,providingmotivationforour modelingapproach. Millaretal.(2016)estimatetime-to-planforcompletedprojectsandhighlight the significant heterogeneity in time from plan-to-start and start-to-completion across projects. SelgradandSiani(2025)demonstratetheinteractionofplannedinvestmentandmonetarypolicyfor firm-levelprojects,demonstratingtheimportantrolenewprojectsplayinrespondingtomonetary policy shocks. Brandsaaset al.(2024) analyzeplans for manufacturingactivity,highlighting how starts rose around the time of the implementation of the Inflation Reduction Act and CHIPS and ScienceAct. Ourworkshowsthevalueofplansinforecastingdatacenterinvestment;theframework cannaturallybeappliedtononresidentialstructuresinvestment,whichtypicallyaveragesabout20% ofprivateinvestment(Brandsaas etal., 2024),or toother formsofinvestment whereabandonments outofplanning,time-to-plan,andtime-to-constructioncanbeconsistentlyestimated. Therestofthepaperfollowsassuch. InSection2,wedescribehowweobtainaggregatedatacenter investmentintheNIPAs. InSection3,wepresenttheframeworkweuseforgeneratingaggregate data center investment from project-level data. In Section 4, we describe the Dodge data and its properties. InSection5,wepresentthesimulationresults. InSection6,weconclude. 2. MEASUREMENTOFDATACENTERINVESTMENTINTHENIPAS TheNIPAsdonotcontainasinglecategoryofinvestmentthatcapturesalldatacenterinvestment. Instead, investment related to data centers is recorded in a few main categories. BEA measures structure investment using the Census’s nonresidential structures CPIP report. The left panel of Figure1showsthatnominalspendingondatacenterstructureshasincreasedrapidly,fromabout $10billiondollarsin2021toabout$40billioninthefirsthalfof2025.3 However, total investment in data centers is much larger than the structures themselves, as data centersareequippedwithhigh-techcapitalgoods,includingservers,storagearrays,andnetworking 3Ourinvestmentsimulationsareinnominalterms,sohereweplotnominalinvestment. Realinvestmentinthese categoriesisshowninSupplementalMaterialsFigureS1. 4

250 200 150 100 50 0 )etar launna( sralloD fo snolliB 250 200 150 100 50 0 2015q1 2017q1 2019q1 2021q1 2023q1 2025q1 High-tech Equipment Trend Structures (a)DataCenterRelatedInvestment )etar launna( sralloD fo snolliB 2015q1 2017q1 2019q1 2021q1 2023q1 2025q1 Imputed Data Center Investment (b)ImputedDataCenterInvestment Figure 1: INVESTMENT IN DATA CENTER CATEGORIES. Notes: Panel (a): Nominal investment for data center structures and high-tech computers & peripheral equipment. The estimated trend is calculatedfrom2015to2022;seeSection2 formoredetails. Investmentindatacenterstructures isonlyavailablefrom2020. Panel(b): Imputeddatacenterinvestment,whichisthesumofdata centerstructuresanddetrendedhigh-techequipment. Source: Authors’calculationsusingdatafromtheBureauofEconomicAnalysis. hardware. Thepurchasesoffinalcomputerhardware,includingserversandotherfinishedcomputing equipment,arecountedinthe“ComputersandPeripheralEquipment”componentofprivatefixed investmentininformationprocessingequipment, shownasthesolid bluelineintheleftpanelof Figure 1. The level of this series also reflects investment in other computer equipment, such as businesspurchasesoflaptopsandprinters. OurapproachistoeffectivelydetrendsuchcomputerE&Iinvestmentandtaketheresidualasthe E&Iinvestmentindatacenters. We estimate atimetrendfrom 2015to2022, shownasthe dotted lineintheleftpanel.4 Thisisanestimateofcounterfactualinvestmentabsenttherecentsurgein datacenteractivity. Next,weinterpretthedifferencebetweenthesolidanddottedlinesasreflecting datacenter equipmentinvestment,andwe addthis differencetostructures investment (black line, leftpanel)toobtainameasureofdatacenterinvestment. This sum is shown in the right panel of Figure 1. Though we think this measure of data center 4Weestimatethetrendthrough2022,topredatethestartoftherampupindatacentersstructuresspendingin2023. Thatsaid,thetrendsaresimilarwhenestimatedthrough2021or2023. 5

investment accurately captures recent trends in data center investment, we note this is still an imperfectestimate. Ontheonehand,itdoesnotincludesomerelevantpieces,suchasinvestmentin specialized data center software or own-account equipment investment.5 On the other hand, this estimate could include some purchases of equipment unrelated to data centers, if, say, business purchasesoflaptopsincreasedmeaningfullyabovetrend. 3. AFRAMEWORKFORFORECASTINGAGGREGATEINVESTMENTWITH PROJECT-LEVELPLANNINGDATA In this section, we describe the framework we use to simulate aggregate data center investment usingproject-leveldata. 3.1. Environment Our analysis is indexed at the project level, and each project i has a vector of characteristics, X. i These characteristics include an estimate of total project cost in dollarsV. The project evolves i throughdifferentstages, j. Theprojectstagescantakeoneoffourvalues: inplanning(P),under constructionor“started”(S),completed(C),orabandoned(A). Allprojectsareassumedtostartwithaplanningstage(P),andthenaftersomenumberofperiods advance to a construction start or are abandoned. Once started, projects are assumed to reach completionaftersomeadditionalnumberofperiods.6 The average time from plan to construction, T (X), can vary with project characteristics, in P,S i particularthesizeoftheproject. Theaveragetimefromstart-to-completion,T (X),cansimilarly S,C i varywithprojectcharacteristics. Theabandonmentrate,λ(X),istheaveragenumberofprojects i thatareabandonedoutofplanningattheonsetofoursimulation,whichcanalsovarywithproject characteristics. Therearevariouspotentialassumptionsonecouldmakewithregardstothedistributionofinvestment across the construction stage—that is, how quickly spending is phased in once construction begins. 5WedonotincludesoftwareinvestmentbecauserecentchangesintotalsoftwareinvestmentlikelyreflectotherAI developmentsbeyondinvestmentsinsoftwaretooperatedatacentersthemselves. Formorecontextonown-account equipmentinvestment,seeByrneetal.(2017)andByrneetal.(2018). Thein-houseassembliesofintermediatecapital purchasesarenotincludedintheequipmentseriesabove. 6Asmallshareofprojectsthatreachconstructionareabandoned—wecanobserveandaccountforthesecasesonce theyoccur—butforestimatingtoplanortime-to-completion,giventheirsmallshare,weassumethiscaseaway. 6

Forouranalysis,weassumethatinvestmentisdistributedevenlyacrosstheconstructiontimeline. In effect, a project with value V and months from start-to-completion of T , spends out I = i S,C i V/T (X) each month from its start date through completion. This assumption would be too i S,C i backloadedifwewereonlyconsideringbuildingconstruction. TheU.S.CensusBureauestimates that for very large projects (> $100 million) that take over 3 years to fully build out, typically about 50 percent of the value is phased-in the first year, and 90 percent by the end of the second year.7 However,itseemsreasonable toassumethatequipmentinstallationwill berelativelymore backloaded than construction of the structure itself. Therefore, this assumption balances these competingforces;nonetheless,thisassumptioncanbeeasilyvariedwithinourmethod. In the simulations and discussion, we find it helpful to distinguish between the “short run” and the“mediumrun.” Intheshortrun—aboutayearfromthelastvintageofplanningdata—thelevel ofinvestmentislargelydeterminedbythestockofcurrent,existingplans,sinceittypicallytakes aboutoneyearforplansto movefromplanningtotheconstructionsstartstage. Whileinvestment intheshortrunislargelyafactorofexistingplans,investmentinthemediumrunisinsteadheavily dependentonassumptionsaboutfutureinflowofplans. 3.2. NowcastingtheShortRun Giventhethreeparametersdefinedaboveandastockofprojectsstillintheplanningstageorwhich arestill underconstruction attimet, wecan generatea nowcast ofthe aggregate timeseries forthe value-in-place fromtheseprojects. The firststep is todenote themontha projectstartsconstruction ass. Forprojectsthathaveobservedstartsinthedata,weknowthestartdate,whileforprojects i thatareplans(thathavenotbeencancelledinthesimulation),s canbederivedfromthemonththe i projectenteredthedataandthetimefromplan-to-startT (X). P,S i Thetotalvalueofinvestmentvalue-in-placeinmontht isgivenbysummingacrossallprojectsin constructioninmontht: I =∑1[s ≤t <s +T (X)]·I, (1) t i i S,C i i i where 1[·] is an indicator function equal to 1 if project i is under construction in month t and 0 otherwise,andI =V/nismonthlyproject-levelinvestmentdefinedasabove. Thus,aprojectthatis i i underconstructionfor,say,twelvemonths(T (X)=12)andstartedconstructioninJanuary2025 S,C i willcontributetoinvestmentthroughtheentireyear. Wenotethatthisframeworkforinvestment— 7Formoreinformation,seehttps://www.census.gov/construction/c30/pdf/t424.pdf. 7

of spending that is phased-in while the project is under construction—is closely related to the methodologyemployedbyU.S.CensusBureautoestimateinvestmentonnonresidentialstructures. However,thisprocessdoesdifferfromhowthecontributionfromequipmentismeasured,which theU.S.CensusBureaumeasuresatthetimeofpurchase. 3.3. ForecastingtheMediumRun In order to extend our simulations beyond one-year ahead, assumptions are required about the the inflowofnewprojectsintoplanning,unliketheshort-termnowcastwhichdependsontheactual announcedstockofprojectplans. Lett¯denotethelastmonthforwhichweobserveplannedprojects,andletT denotetheaverage P,S timefromplanningtoconstructionstart acrossallcharacteristics. Newprojectsenteringplanning aftert¯willnotmeaningfullycontributetoinvestmentuntilmontht¯+T . Wedefinethestartofthe P,S mediumrunas: k¯ =t¯+T (2) P,S Forexample,ift¯isDecember2025andT is10months,thenk¯ isSeptember2026. Inpractice, P,S investment in monthk¯ will bedominated by projectsfrom the existingstock with start dates s <k¯. i Astheforecastperiodincreases,theshareofinvestmentattributabletonewprojectinflowsincreases. 4. PROJECT-LEVELDATA This section first describes the project-level data we use to generate aggregate measures of data center investment. We then present descriptive statistics from our data sample. Last, we provide detailonthesampleusedinoursimulationexercise. 4.1. DodgeConstructionNetworkDataonDataCenters OuranalysisisbasedonRealEstateAnalyzer(REA)datafromDodgeConstructionNetworkfrom 2003toSeptember2025. Dodgeisaleadingdataproviderintheconstructionindustryandtheirdata areusedbytheU.S.CensusBureautomeasureCPIP.Specifically,theCensussurveysastratified sampleofconstructionprojectsobtainedfromDodge,andprojectmanagersthenreportonthevalue ofwork(“value-in-place”)doneeachmonthfromprojectstart-to-completion. Previouswork,for 8

instanceBrandsaasetal.(2023),hasshownthatdataonDodgeconstructionstartsarehelpfulfor predictingfuturespendingonstructures. In our work here, we instead mainly focus on construction project plans rather than starts for a few key reasons. First, plans increase the lead time for economic projections, because they lead constructionstartsbyaboutayearonaverage. Second,intheREAdata,thedollarvalueofprojectlevelplanslikelymainlyreflectstotalcostsfortheproject(includingbothequipmentandstructure). We verify this is the case by comparing plan costs in REA versus costs we found online for the largestrecentlyannounceddatacenterprojects. Theproject-levelplansareespeciallyvaluablein thiscontextsincewecareabouttotalinvestmentratherthanstructuresinvestmentalone.8 Basing our approach on actual project level plans is thus useful for estimating and nowcasting overall investment in data centers related to both equipment and structure as measured in the NIPAs (as describedinSection2).9 TheREAdatacontaininformationoneachproject’spropertytype,squarefootage,location,and project value. The total project values we use are in nominal US dollars and we only consider projectsbeingbuiltintheUnitedStates. Wetakethelastreportedplanvaluetobethevalueofthe plan. Weconsolidateprojectstagesintofourcategories: planning,start,completed,andabandoned. TheanalysisrequiresharmonizingtheREAdatainordertoconsistentlytrackprojectsfromplanning tosubsequentphases. Forinstance,weassessaprojecttohaveenteredthestartstagethelasttimeit doessowithout becomingaplanagain. Wealso assumeabandonsandcompletionscan onlyoccur at theend ofa project’s series. Allpossible phasechanges areshown inFigure S4and information onhowthesephasechangesarederivedcanbefoundinSectionS.2.1,bothintheSupplementary Materials. Somelargeprojectsareinitiallyreportedasmasterreportsandthenreceivechildreports forindividualstructuresorbuildingphases.10 Dodgeprovidesamappingbetweenmasterandchild 8IntheREAdata,project-levelvaluescanbeupdatedovertime,andinoursimulationsweusethelastavailable estimateofprojectvalue(asofSeptember2025). GivenDodge’sfocusontheconstructionindustry,itispossiblethat someproject-levelcostsrevisedowntohoneinonstructurecosts,astheprojectmovesfromplanningtoconstruction stage. Whileprojectvaluesdorevise,theydonotshowastrongtendencytorevisedown(orup)astheymovethrough constructionstages. Hence,ourinterpretationthatprojectvaluesmainlyreflecttotalprojectcosts. 9Industryanalystshaveimputedthelevelofdatacenterinvestmentusingotherapproaches. Forinstance,analysts atJ.P.Morgan(ReinhartandFeroli,2025)forecastinvestmentbasedonestimatesofannounceddatacenterpower consumptionandassumptionsonthecostsofachievingthoseplannedgoals. AnalystsatGoldmanSachs(Pengetal., 2025)generateestimatesusingchangesinrevenueorreportedinvestmentincompaniesorindustriesexposedtotheAI boom. 10Tofocusondatacenters,wesubsetthedatatoallprojectswhosemostrecentlyrecordedstructuretypeisadata center,aswellasthemasterreportsfortheseprojectsregardlessoftheirstructuretype. 9

400 300 200 100 0 sralloD fo snoilliB 1500 1000 500 0 2022 2023 2024 2025 2026 (a)ValueofNewPlansoverTime sralloD fo snoilliB 2022 2023 2024 2025 2026 (b)PlanningStockValueoverTime Figure2: VALUE OF NEW PLANS AND PLANNING STOCK OVER TIME. Notes: Panel(a)shows thevalue ofnew datacenter plansina givenmonth. Panel(b) showsthe totalstock ofdatacenter projectscurrentlyintheplanningstageinagivenmonth. Source: Authors’calculationsusingdatafromDodgeConstructionNetwork. reports; our process for linking these reports is outlined in the Supplementary Materials in Section S.2.2. 4.2. DescriptiveStatistics TheleftpanelofFigure2plotsmonthlyunadjustednominalvalueofnewdatacenterplansinthe Dodge data.11 Prior to 2023, data center plans were modest in scale; for instance, they averaged about$7 billionper year from2019 to 2022. However, data centerplans began toskyrocket inlate 2023,cumulating tonewplanvaluesof almost$300billionin 2024andover$900 billionin2025 (asofSeptember). Assomeoftheseplansenteredtheconstructionphase,theycontributedtogreater data center spending, as shown in Figure 1. However, most of these plans have not yet broken ground. TherightpanelofFigure2showsthatthevalueofthestockofdatacenterplans—which cumulates the inflow of plans while excluding plans that have begun construction or have been abandoned—exceedsover$1trilliondollarsasofSeptember2025. Theincreaseinplansreflectsbothanincreaseinthenumberofannounceddatacenterprojects,as 11Wechoosetopresentnominalspending(ratherthanrealspending),asforecastingdeflatorsforspendingondata centerstructuresandequipmentwouldintroduceanotherlayerofuncertainty. 10

well as an increase in the dollar value of the plans. For instance, in the four years from 2019 to 2022, the data includes plans for about 278 data centers, with an average project value of about $114millionandmedianvalueof$26million. In2025throughSeptember,therewere521plans, withanaveragevalueofalittleunder$2billionandmedianvalueofabout$216million. 4.3. Estimationsample AsdiscussedinSection3,tosimulateinvestmentwerequireestimatesoftimefromplan-to-start, time from start-to-completion, and abandonment rates. To obtain these estimates, we focus on a sampleofprojectswithobservedplanningstagesandwithobservedtransitionsoutofplanning.12 Thisdropsmanyrecentplansfromtheestimationsample(sinceitistooearlytodeterminewhether they will build out or not), but we include these later in the simulation sample. We also exclude master reports, which aggregate multiple projects and have ambiguous abandonment status, and theseprojectsarealsoreintroducedinthesimulationsample. Intotal,thesereductionsreducethe estimationsamplesizeby56%,predominantlydue torecentprojectsthathavenottransitionedout ofplanning. Table1belowshowssummarystatisticsforabandonmentrates,timeinmonthsfromplan-to-start andstart-to-completion,andprojectvaluesforallprojectsintheestimationsampleinpanelA,and forrelativelylargeprojects—thoseworthover$250milliondollars,correspondingroughlytothe topdecileofplans—inpanelB.StartingwithpanelA, which isbasedonabout1,700projects,we estimateanabandonmentrateof0.33. Meanwhile,ittakesabout11monthsforprojectstomove fromplan-to-start,andonaverage,about9monthsforprojectstomovefromstart-to-completion,as recorded in the REA data. The statistics on months of time from plan-to-start, and from start-tocompletion, aresimilar tothose reported inGlancy et al.(2025) forthe overall commercialsector. However,theabandonmentrateislowerthanthe0.46abandonmentratereportedinGlancyetal. (2025), suggesting data centers have somewhat lower abandonment rates, typically, than typical officeorretailbuildings. Turning to panel B, for the relatively larger projects the abandonment rate is about 0.3, similar to the rest of the sample. Consistent with Glancy et al. (2025), we find that abandonment rates do not vary much (unconditionally) by size, perhaps reflecting countervailing forces. On the 12Weclassifyprojectsasabandonedoncetheyhavebeenobservedinplanningforatleast48withoutanobservedstart phase,sinceabout95%ofprojectsthatbeginconstructiondosowithinfouryears,seeFigureS2intheSupplemental Materials. 11

Mean S.D. p10 p50 p90 Count PanelA:Allprojects Fractionabandoned(λ) 0.33 0.47 0.00 0.00 1.00 1711 Monthsplantostart(T ) 11.11 15.36 2.00 6.00 26.00 981 P,S Monthsstarttocompl(T ) 9.37 6.83 3.00 9.00 17.00 715 S,C Valueinbillions(V) 0.10 0.38 0.00 0.01 0.25 1711 PanelB:Largeplans(over250millionUSD) Fractionabandoned(λ) 0.30 0.46 0.00 0.00 1.00 169 Monthsplantostart(T ) 15.00 18.47 2.00 9.00 37.00 113 P,S Monthsstarttocompl(T ) 19.77 9.78 10.00 18.00 35.00 48 S,C Valueinbillions(V) 0.70 1.02 0.28 0.48 1.07 169 Table1: ESTIMATION SAMPLE SUMMARY STATISTICS. Notes: Thetableshowssummarystatistics (mean,standarddeviation,percentilesofthedistribution,andcount)fordatacenterprojectsinthe estimationsample. Monthsfromplan-to-startandstart-to-completion(compl)areonlycalculated forprojectswithstartandcompletionentries,respectively. Source: Authors’calculationsusingdatafromDodgeConstructionNetwork. one hand, abandonment rates for large projects may be more sensitive to changes in economic conditions, as suggested by Glancy et al. (2025). On the other hand, the firms involved in larger projects are likely to differ—for example, they potentially have easier access to credit. Turning to construction timelines, we find months from plan-to-start of about 15 months for these larger projects,alittlehigherthanthe11monthsintheoverallsample(notstatisticallydifferent). Thereis amoremeaningfuldifferenceinmonthsfromstart-to-completion(20vs9months),showingthat, asexpected,largerprojectstakelongertoconstruct. 5. DATACENTERINVESTMENTSIMULATION In this section, we present the results from our simulation exercise. We first present the calibration details; we then present results for the short run and medium run, with the latter requiring assumptionsabouttheinflowofnewplansgoingforward. 12

5.1. Calibration Wecalibratetheparameterdistributionsbasedontheestimationsampleasfollows. Forthedistributionoftheabandonmentrateλ,wesetthemeanto0.33andthestandarddeviation tothestandarderrorofthemean0.01. Forthedurationofmonthsfromplan-to-startT ,wesetthe P,S meanto11.11andthestandarddeviationto0.49.13 Becausetimefromstart-to-completionT (X)hasastrongassociationwithprojectsize,wemodel S,C i itasfollows: T (X)=β +β ·log(V ). (3) S,C i 0 1 i Thisregressionyieldsβˆ =19.15(SE=0.54)andβˆ =2.15(SE=0.11). Thisequationimpliesa 0 1 $1billiondatacenterprojecttypicallytakesabout19monthsfromstart-to-completion. The discussion above describes our treatment of plans. For projects that we see in the data that have already begun construction, we simulate spending based on the observed start date and the time-to-completioncalibrationapproachdescribedabove. 5.2. Short-RunNowcasts The sample used in the simulation includes all projects that are in planning or that have started butarenotcompletedorabandonedinthedatasetprovidedbyDodgeinSeptember2025. Inthe simulations,there aretwobroad sourcesofrandomness. First,ineach simulation keyparameters areindependentlydrawnfromnormaldistributions,asdescribedintheprevioussubsection. Second, in each simulation, abandonments are determined stochastically at the project level. In some simulations,many ofthe verylargeprojects endup abandoned, and, asaresult, totalinvestmentis relativelylower,andviceversa. Inpractice,mostoftherandomnesscomesfromtheproject-specific abandonment. Togeneratenowcasts,weusethemethoddescribedinSection3. Wegenerate1000MonteCarlo simulationsofinvestmentovertheshortrun. Afterdrawingproject-levelabandonments,weobtain a set of non-abandoned projects. Investment is then simulated for these projects based on the time-to-buildparametersandtheassumptionsdescribedabove. The leftpanel ofFigure 3 presentsthe simulationresults forthe short-run periodthrough 2026:Q2. 13ThemeansarereportedinthefirstcolumnofTable1. Thestandarderrorofthemeanisgivenbytheratioofthe standarddeviation(secondcolumn)andthesquarerootofthenumberofobservations(lastcolumn). 13

The figure displays data center investment at a quarterly frequency (annualized) across various distribution points: mean, median, interquartile range, and additional percentiles. In the mean simulation, investment increases from approximately $60 billion in 2024:Q4 to $180 billion in 2025:Q4,reachinganannualrateofroughly$370billionby2026:Q2. As avalidation exercise,we compare simulatednominal investmentto theNIPA-based imputation of data center investment discussed in Section 2. This estimate is shown as the red dotted line in theleftpanelofFigure3,andisthesumofnominaldatacenterstructureswith“excess”high-tech equipment,definedasthedifferencebetweenmeasuredinvestmentinthecomputersandperipheral equipmentcategoryandatrendestimatedfrom2015to2022. Inthefigure,theredlineissimilar tosimulatedinvestment,indicatingthat,thusfar,simulatedinvestmentisagoodguideforactual investment. Thesimulationsexhibitsubstantialdispersion. Atthe5th percentile,investmentreachesapproximately$240billionin2026:Q2,whileinvestmentstandsatabout$450billionatthe95thpercentile. Thisvariationprimarilystemsfromproject-levelabandonmentrates.14 Insomesimulations,many recentlyannouncedlargedatacentersarecompleted,whileinothers,manyabandonsoccur. The median exceeds the mean in these simulations because the distribution of total investment, shown in Figure S3 in the Supplemental Materials, displays bimodality with greater mass in the right mode. This distribution pattern results from two factors: the abandonment rate of 0.33, and the right-skewed distributionof project-level investment values. Specifically,a smallnumber of extremely large projects, such as Project Stargate, represent significant outliers. Simulations where these outsized projects are abandoned yield lower overall investment levels. Conversely, when completed,they substantiallyelevate investment totals. Since projectssurvive approximately two-thirdsofthetimeinthemodel,theright-sidedistributionmodecontainsgreatermass. Hence, most simulations include the completion of the very large projects, thereby pushing the median abovethemean. WeinterpretthisnowcastasconsistentwiththeGabaix(2011)“granularorigins”viewofaggregate fluctuations. The distribution of data center projects has a fat right tail, so the magnitude of the investment boom depends critically on the spending and completion of a handful of very large 14TheconfidenceintervalsstartpriortothelatestvintageofDodgedata,becauseinthesimulations,projectscan break ground before they enter the construction stage in the Dodge data. As an example, a project that enters the planningstagein2024:Q1couldbreakgroundinsomesimulationslaterintheyeareveniftheDodgedatadonotyet indicatetheprojecthasbegun. 14

500 400 300 200 100 0 )etar launna( sralloD fo snolliB 2023q1 2024q1 2025q1 2026q1 Mean Median 25-75 5-95 1-99 Min-Max Data (a)ShortRun 1000 800 600 400 200 0 )etar launna( sralloD fo snolliB 2025q1 2025q3 2026q1 2026q3 2027q1 2027q3 Quarter Short-run Inflow 2x Inflow 1x Inflow 0.25x (b)MediumRun Figure3: DATA CENTER INVESTMENT. Notes: Panel(a)displayssimulatedshort-runinvestment based on 1000 simulations. The red dotted line represents data center investment imputed from NIPAdata(seeSection2). Inpanel(b),wesimulatethrough2027basedonthreedifferentscenarios forthelevelofnewplansinthefuturebasedonthelagged-levelofnewplanstimesascalingfactor (2ingreen,1inyellow,or0.25inred). Theshadedareasdenotethe5th-95thpercentiles. Source: Authors’calculationsusingdatafromDodgeConstructionNetwork. 15

investment projects. That said, it is worth noting that the rise in data center projects has been sufficiently broad-based that, even in the most pessimistic scenarios where many of these large projectsarecompletelyabandoned,investmentstillrisesto$240billionbythemiddleof2026. 5.3. Medium-RunForecasts The simulations through the short run are independent of our assumptions about the future inflow of plans, as plans take about a year in the simulations before they break ground. We now turn to simulating investment through 2027 (which we label as the medium run), which requires assumptions about the inflow of new plans through 2027. We first calculate the average plan size and the monthly flow of new plans from 2024 to September 2025. We then consider three alternatives forthenumberof plansgoingforward toprovidearange ofscenarios: a)one-fourth, b) the same,or c) twicethe numberof plans permonth inthis recenthistory, holding the project size constant. Lastly, wesimulateinvestment through2027accordingtothesethreedifferent scenarios; the results are shown in the bottom panel of Figure 3 and the first three columns of Table 2. The figureincludesthe5thand95thpercentilesforeachsimulation. Thesimulationsstarttobranchoutinlate2026,whentheassumedplansinlate2025startbreaking ground. Bytheendof2027,investmentdiffersmarkedlydependingontheassumptionsaboutplans, withmeaninvestmentrangingfromabout$360billioninthemorepessimisticalternativeto$930 billionin the moreoptimistic scenario. Investmentis relativelyrobust, even inthe scenariowhere futureplansfalltoaquarteroftheirrecentlevel,asdatacenterprojectscurrentlyinplanningsupport investmentforafewyears. Hence,the2027forecastsreflectamixofbothcurrent“data”—plans thathavealreadybeenannounced—aswellasthesimulatedpathsofplansgoingforward. Looking beyond2027,simulatedinvestmentwouldrelyevenmoreheavilyontheassumptionsaboutnew plans. Table 2 also reports the data center investment contributions to nominal GDP growth from the simulations. Tocalculatethecontributions,wegrowoutnominalGDPin2024usingnominalGDP growthforecasts available from the Survey of Professional Forecasters(SPF).15 In the middlethree columns, we report these contributions before accounting for the share of high-tech equipment that is imported. In the mean simulation, data center investment contributes 0.4 percentage points to 15TheSPFdoesnotincludea2027nominalGDPforecast. Toobtainit,weusetherealGDPgrowthforecastandan extrapolationoftheGDPpriceindexfrompreviousvalues. Sincetheseforecastslikelyincorporatesomeboostfrom datacentergrowth,wedonotadjustnominalGDPinthethreedifferentscenarios. 16

Investment(Bill. $) ContributionstoGDPGrowth(p.p.) IncludingImportedInputs ExcludingImportedInputs Year 0.25x 1x 2x 0.25x 1x 2x 0.25x 1x 2x 2023 10 10 10 0.02 0.02 0.02 0.01 0.01 0.01 2024 58 58 58 0.17 0.17 0.17 0.07 0.07 0.07 2025 179 179 179 0.41 0.41 0.41 0.18 0.18 0.18 2026 413 464 526 0.75 0.91 1.11 0.33 0.40 0.49 2027 357 612 921 -0.17 0.45 1.21 -0.08 0.20 0.53 Table 2: DATA CENTER INVESTMENT & CONTRIBUTION TO GDP GROWTH BY INFLOW OF FUTURE PLANS. Notes: The table shows mean data center investment in the fourth quarter (annual rate) and investmentcontributionstoGDPgrowthbasedondifferentassumptionsaboutthefutureinflowof plansasdescribedinSection5.3. TheinvestmentcontributioniscalculatedastheshareofGDPin thepreviousyeartimestheQ4/Q4growthrateofdatacenterinvestment. Ourimportadjustment assumes that 44% of investment is domestic; see Section 5.3 for details. Nominal GDP growth comes from the Survey of Professional Forecasters (SPF). Because the SPF does not include a 2027nominalGDPforecast,weusetherealGDPgrowthforecastandanextrapolationoftheGDP priceindexfrompreviousvalues. SPFdataisavailablethroughtheResearchDepartment,Federal ReserveBankofPhiladelphia,athttps://www.phil.frb.org/research-and-data/real-time-center/surveyof-professional-forecasters. Source: Authors’calculationsusingdatafromDodgeConstructionNetworkandtheSPF. GDP growth in 2025. In 2026, the contributions range from about 0.8 to 1.1 percentage points, depending on the assumptions about future plans. Finally, in 2027, the contributions to GDP growthdiffermarkedly,rangingfrom1.2percentagepointsinthemoreoptimisticscenarioto-0.2 percentagepointsinthemorepessimisticone. However,theneteffectofthedatacenterinvestmentboomonGDPgrowthwillnotonlydependon thelevelofinvestment,themainfocusofouranalysis,butontheshareofinvestmentnetofimports. Inturn,werequireanestimateoftheimportshareofdatacenterinvestment;wecreatea“ballpark” estimate in the following way. First, we assume that 30 percent of investment is in domestic structures and the remaining 70 percent in high-tech equipment. We derive this split using BEA andCensusestimatesofdata centerstructurevalueputinplaceandhigh-techandother equipment spendingbyinformationanddataprocessingfirmsin2024. Second,weassumeanimportshareof 80% for high tech equipment. The import share is based on imports and total investment in high 17

tech equipment from BEA and Census, also in 2024. Combining these assumptions implies that 44%ofdata centerinvestment isdomestic. Theresultingcontributionsare shown inthefinalthree columnsofTable2,andarejust44%ofthemiddlethreecolumns. Importantly,theimport-adjustmentaddsanotherlevelofimprecisiontoourforecasts. Anadvantage ofourmethodintheshortrunisthatitgeneratesnowcastsforthetotallevelofdatacenterinvestment (and hence total GDP contributions) with very little judgement, as it is largely pinned down by thecurrentstockofplansandhistoricalestimatesforthemodelparameters. However,accounting for imports requires more judgement about the import shares in coming quarters. The medium run forecasts for data centers’ contribution to GDP growth require assumptions on both import sharesandtheinflowofplans;giventhatbothmarginsarecurrentlyunknown,thereissubstantial uncertaintyaroundanyforecastofsuchinvestmentinthemediumrun. 5.4. FurtherRobustnessandDiscussion Webrieflydiscusshowchangestokeyparameterswouldaffectthesimulations. Theabandonment rate of 0.33 was chosen from the historical average, but the yearly abandonment rates vary from about 0.2 to 0.45. Changes in the abandonment rate can have subtle effects on the percentiles of the distribution of simulated investment. For example, if abandonment rates rise above 0.5, the distributionofinvestmentwouldremainbimodal,butwithgreatermassintheleftmode,triggeringa downwardshift inthemedian. However,themeanscales roughly proportionallywithabandonment rates. Forinstance,achangeintheabandonmentratefrom0.33to0.45wouldroughlyimplyscaling down mean investment levels by 18 percent (as the survival rate shifts down 18 percent from 67 to55percent). Onemightexpecttheabandonmentrateparametertoincreaseiftheexpectations around the potential returns to AI were to worsen. One advantage of the bottom-up simulation approachdevelopedinthispaperisthatitcandynamicallyadjustassuchabandonmentsoccur. A related discussionpertains toscaling factorsover thelifetimeof theproject; that is,it ispossible thatmanyoftheprojectssurviveandbuildout,butareeventuallyscaleddownorupinscope. Inthe historical datacenter data, wedid notsee a strongtendency forvalues to scaleone wayor another, and so we opted to leave scaling factors out of the baseline. With a different assumption, mean investmentwouldscaleproportionately,solongasthescalingfactorappliesequallytoallprojects. It also seems plausible that the unprecedented acceleration in data center construction could be delayed by production capacity constraints. While predicting when and where supply chain 18

bottlenecksmanifestisnotoriouslychallenging,therearesignsofstrainacrossmanyoftheinputs neededtobuilddatacenters,includingland,labor,equipment,power,andcapital(see,e.g.,Chen (2025),TurnerandWest(2025),DuguidandKinder(2025),Whelan(2025)). Thesimulationscould beadjusted with differentassumptionsabouttime fromplan-to-startor start-to-completion. Shifts intheconstructiontimelinewouldroughlydelaysimulatedinvestmentproportionally. 6. CONCLUSION We provide a new method to generate real-time estimates and short-run nowcasts of aggregate spendingondatacenterinvestmentbasedonproject-leveldatausingasmallnumberofempirically estimated parameters. This method can also generate medium-run forecasts with assumptions on theinflowoffutureplans. Ourestimatesperformswellout-of-sampleinto2025,andsoweuseit tonowcastinvestment in2026 givenobserved planningand constructionactivityas ofSeptember 2025. We show how to use our methodology to forecast further out as well, but these mediumrunforecastsbecomequitesensitivetoassumptionsaboutinflowsofplanningactivitywhichare currentlyunknown. Though data center investment is of particular interest given the large investments in the space duringthepost-pandemicera,ourmethodologycanbevaluableforexamininginvestmentofany category duringpotentialinflection points,as longas comprehensive project-level microdataare available. Thatsaid,oursimulationsdonottakeintoaccountthevastarrayofothermacroeconomic effectsfromthesurgeindatacenterinvestmentortheadoptionofartificialintelligence. Separating outthesechannelsastheeconomyevolveswillbeanimportantchallengeforfutureresearch. References Acemoglu,Daron,“TheSimpleMacroeconomicsofAI,”EconomicPolicy,2025,40(121),13–58. Brandsaas, Eirik E., Daniel I. Garc´ıa, Joseph B. Nichols, and Kyra Sadovi, “Nonresidential ConstructionSpendingisLikelyNotasWeakasitSeems,”FEDSNotes.Washington: Boardof GovernorsoftheFederalReserveSystem,2023. Brandsaas, Eirik E., Robert J. Kurtzman, and Joseph B. Nichols, “From Plans to Starts: Examining Recent Trends in Manufacturing Plant Construction,” FEDS Notes. Washington: BoardofGovernorsoftheFederalReserveSystem,2024. 19

Brynjolfsson, Erik, Danielle Li, and Lindsey Raymond, “Generative AI at Work,” The Quarterly JournalofEconomics,May2025,140,889–942. Byrne,DavidM.,CarolCorrado,andDanielE.Sichel,“Own-AccountITEquipmentInvestment,” FEDSNotes.Washington: BoardofGovernorsoftheFederalReserveSystem,2017. Byrne,DavidM.,CarolCorrado,andDanielE.Sichel,“TheRiseofCloudComputing: Minding YourP’s,Q’sandK’s,”NBERWorkingPaperNo.25188,2018. Chen, Te-Ping, “Data Centers are a ‘Gold Rush’ for Construction Workers,” Wall Street Journal, https://www.wsj.com/business/ data-centers-are-a-gold-rush-for-construction-workers-6e3c5ce0. November292025. Duguid, Kate and Tabby Kinder, “Investor Angst Over Big Tech’s AI Spending Spills Over into Bond Market,” Financial Times, https://www.ft.com/content/ d2bf6c25-fb42-4f13-b81c-a72883632f50.November112025. Furman,JasonandRobertSeamans,“AIandtheEconomy,”InnovationPolicyandtheEconomy, January2019,19,161–191. Gabaix,Xavier,“TheGranularOriginsofAggregateFluctuations,”Econometrica,May2011,79 (3),733–772. Glancy, David, Robert J. Kurtzman, and Lara Loewenstein, “JUE Insight: Shovel Ready Projects and Commercial Construction Activity’s Long and Variable Lags,” Journal of Urban Economics,2025. Millar,JonathanN.,StephenD.Oliner,andDanielE.Sichel,“Time-to-PlanLagsforCommercial ConstructionProjects,”RegionalScienceandUrbanEconomics,2016,59,75–89. Peng, Elsie, Joseph Briggs, and Sarah Dong, “US Daily: A Primer on the Impact of AI on the GDPStatistics,”GoldmanSachsEconomicResearch,September132025. Reinhart,AbielandMichealFeroli,“US:TheDataonDataCenters,”J.P.Morgan: NorthAmerica EconomicResearch,January152025. Selgrad,JuliaandKerrySiani,“MonetaryPolicyandInvestmentPlans,”WorkingPaper,2025. 20

Stock,JamesH.andMarkW.Watson,“EvidenceonStructuralInstabilityinMacroeconomic TimeSeriesRelations,”JournalofBusiness&EconomicStatistics,1996,14(1),11–30. Turner, Nicol and Darrell West, “The Future of Data Centers,” Brookings, https://www. brookings.edu/articles/the-future-of-data-centers/.November52025. Whelan, Robbie, “CoreWeave Reports Doubling of Revenue From AI Boom,” Wall Street Journal, https://www.wsj.com/business/earnings/ coreweave-earnings-q3-2025-crwv-stock-cd0f9c8c?.November10,2025. 21

SUPPLEMENTARYMATERIALS This document contains the supplementary materials as referenced in the manuscript. The first section contains the supplementary figures that are referenced in the text. The second section provides further detail on how we clean and process the data from Dodge Construction Network (Dodge). S.1. SupplementaryFiguresReferencedinText 250 200 150 100 50 0 )etar launna( sralloD 7102 fo snoilliB 2015q1 2017q1 2019q1 2021q1 2023q1 2025q1 High-tech Equipment Structures FigureS1: INVESTMENT IN DATA CENTER CATEGORIES. Notes: Realinvestmentby quarterfor datacenterstructuresandhigh-techcomputers&peripheralequipment. Source: Authors’calculationsusingdatafromtheBureauofEconomicAnalysis. 1

100 80 60 40 20 0 detratS eulaV tcejorP fo % evitalumuC Time to Start Distribution (Value-Weighted) 0 50 100 150 Months from Plan to Start Figure S2: TIME-TO-START CDF. Notes: Distribution of months from plan-to-start across all projectswithobservedplanandstartphases. Source: Authors’calculationsusingdatafromDodgeConstructionNetwork. .005 .004 .003 .002 .001 0 ytisneD Distribution of Total Simulated Investment 1 .8 .6 .4 .2 0 400 600 800 1000 1200 Total Investment Value (Billions USD) (a)Density ytilibaborP evitalumuC CDF of Total Simulated Investment 400 600 800 1000 1200 Total Investment Value (Billions USD) (b)CDF FigureS3: DISTRIBUTION OF SHORT-RUN SIMULATED INVESTMENT VOLUME. Notes: Values are based on 1000 simulations of short-run investment. The vertical black lines denote the mean (solid)andmedian(dashed). Source: Authors’calculationsusingdatafromDodgeConstructionNetwork. 2

S.2. DataConstructionDetails This sectiondiscusses how weharmonize the project-levelDodge data sothat construction phases are defined consistently and transitions between phases conform to a consistent structure. This section also discusses how we integrate master reports and their linked child reports, as well as otherdataadjustments. S.2.1. PhaseWork IntheoriginalDodge reports,projectscanmoveinandoutof anyoftheeight phasesdesignated byDodge: pre-planning,planning,finalplanning,bidding,underway,completed,abandoned,and deferred. We make adjustments to create a consistent phase structure across projects. First, we combine pre-planning, planning, final planning, and bidding into a single ”planning” phase. We thenmakephaseadjustmentssuchthatonlyphasechangesshowninFigureS4arepermissible. We achievethisstructurebymakingthefollowingadjustmentstothedatainsequentialorder: 1. Foreachproject,thelasttransitionintothestartphaseisfound. Allprecedingentrieshave theirphasechangedtoplan. 2. Foreachproject,thelasttransitionintotheplanphaseisfound. Allprecedingentrieshave theirphasechangedtoplan. 3. Foreachproject,anyseriesofabandons/deferralsthatdonotcomeattheendoftheproject areswitched toeitherplanorstart, dependingonwhich ofthose twophases mostrecently precededthestringofabandons/deferrals. Additionally,anyprojectthatisintheplanning stagefor48monthsandtheirlaststagelabelisabandonedorplan,hastheirphasechangedto abandoned,startingatthe48+1monthmark. Thisisbecauseabout94%ofprojects(weighted byprojectvalue)withobservedstartdatesstartwithin48months(seeFigureS2). 4. Completionscanonlycomeattheendoftheseriesofreportsbecauseofhowthecompletion variable is defined: every entry after the last updated completion date is considered a completion. Takentogether,thesechangesensurethefollowing: 1. Abandoned or completed entries can only occur at the end of the series of entries for a project. Aprojectcannolongerbelabeleddeferred. Ifaprojectisongoingandiscurrently deferred,itislabeledasabandoned. 3

Entry Plan Start Complete Cancel FigureS4: VALID PHASE TRANSITIONS. Note: Solidarrowsindicateallowabletransitions. The dashed arrows from entry represents the conceptual point at which a project enters the database, ratherthananobservedphase. 2. Onceaprojectmovesoutofaplan,itwillneverbecomeaplanagain. Onceaprojectmoves outofastartitwillnevermovebackintoastart(seearrowsinFigureS4). 3. Projectscanenterthedatasetatanyofthefourredefinedphases. S.2.2. MergingMasterReportswithChildReports Larger projects are often first reported as master reports and, when additional data are available, childreportsaremadeforspecificcomponentsorphasesoftheproject. Childreportsreceiveunique project identifiers and Dodge provides a crosswalk linking master reports with their child report, oncethechildreporthasreceiveditsfirstentry. Wemergechildreportswiththeirmasterreportinthefollowingway. Forallentries,wetakethe mostrecentvalueofthemasterreportasthevalueoftheprojectandderivechangestophasesfrom the masterand child reports. We consider aproject to have started constructionwhen the first child report has started construction. A project is completed when all the child projects complete, or whenatleastonechildprojecthasbeencompletedandallotherchildreportshavebeenabandoned (orhaveapendingdeferral). Weconsideraprojecttohavebeenabandonedwhenallchildreports havebeenabandoned. Theresultingdatastructureissuchthateverymaster-childgroupinghasonly oneentryforeverydatewheredatawasavailableforthemasterreportoroneofthechildreports. We removed one project that matched with multiple master reports and its corresponding master reportsfromthedata. Thereweresixteenprojectsforwhichachildreportwasmadebeforethefirst parentreport. Inthirteen ofthesecases, theprecedingchildreports wereallplans, intwocases a 4

startreporthadbeen madeforthechildprojectbefore thefirstmasterprojectreportwasmade, and in one case a completion reportwasmade for the child project before the first master report entry. Inallsixteencases,wetreatthehistoryoftheprojectashavingstartedatthedateofthefirstmaster projectreportentryandtakethephasevaluesfromthechildprojectsmovingforward. S.2.3. ProjectRemoval After a project has been reported on by Dodge, it may be removed from the set of projects that Dodgetracksforseveralreasons,includingbutnotlimitedto: 1. Thereportisaduplicatereportofanotherproject. 2. Theproject’svaluefellbelow500thousandnominalUSdollars. 3. Theprojectwasamasterreportandallofitschildprojectshavereceivedtheirfirstreports. 4. The projecthas not receiveda phase updatein thirty-six months. Should theproject later be updated,itwillbere-introducedbyDodgetothedata. Dodge provides a removal date, which reflects the date at which Dodge will no longer provide updates to the project. Dodge also provides a comment on the reason for removal. We remove projects from the sample as of their removal dates, unless they are duplicate reports, in which case weremoveallentriesforthatprojectfromthedataset. Separately,weretainonlyprojectreportslocatedintheUnitedStates. Once the data cleaning described above is complete, we have a data set where we observe 3916 projectsovertime,witheachprojectobservedfor58monthsonaverage,withnomissingperiods. Allmasterandchildreportgroupingsareconsolidatedintoasingleseriesofprojectentries. 5