2026全球AI數(shù)據(jù)中心建設競速研究報告第一部分

STUDY

Theglobal

build-outrace

AIdatacenterstudy|Part1

MANAGEMENTSUMMARY

Theglobalbuild-outrace

AIdatacenterstudy|Part1

DatacentersusedtobeanichetopicinenterpriseIT.Today,theysitattheheartofaglobalraceto

build,control,andleveragethecomputational

powerthatwilldefinetheartificialintelligence(AI)erawithconsequencesforGDPgrowth,economicinfluence,nationalsovereignty,andgeopolitics.

ThesefacilitiesaretheveryfoundationoftheAI

revolutionignitedbytheemergenceofgenerativeAIin2022withfar-reachingimplicationsforindividuals,industries,andbusinessesalike.ThepredictedvaluegenerationpotentialofAIhasattractedcapitalatanunprecedentedpace.By2026,hyperscalercapital

expenditure(CapEx)issettoapproachUSD750billion

fuelingtremendousgrowthalongeverylayerofthedatacentervaluechain.

Twostructuralforcesaredrivingdemand:the

exponentialcomputerequirementsofsuccessive

largelanguagemodel(LLM)generations,andan

adoptioncurvewithoutprecedent.TheRolandBergerDatacenterModelprojectsglobalinstalledcapacitygrowingfromroughly90gigawatts(GW)in2025to

nearly350GWby2035,withAI'sshareoftotalloadrisingfrom20%tonearly60%.

Thisgrowthstoryisstrongandfundamentally

soundyetfinancialdynamicsandotherstructuralrisksintroducevulnerabilitiesthatmeritscrutiny.

BothhyperscalersandAImodeldevelopersare

currentlyinvestingmorethantheirAI-linkedrevenues

support,withsomehyperscalers'overallfreecash

flowtightening.Thestronggrowthalongthesupplychainalsocreatestheconditionsforapotential

bullwhipcorrection,andtheriskofsomeshort-term,temporaryslowdowntriggeredbyatighteningof

externalfunding,regulatoryheadwinds,orsupply

chainbottlenecksisnon-negligible.Ecosystem

participantsthatassumearesilientpositionacrossthestackwillbethosethatcapturesustainablelong-termvaluethrougheventualcycles.

Geographically,theUSisatleastmaintainingits

leadinglobalcapacity,reflectingthelocationof

hyperscalersandfrontierlabsandamoreenabling

regulatoryenvironment.Europe'sshareisanticipatedtodecline,constrainedbygridconnectionleadtimes,highenergycosts,andregulatorycomplexity.Easingthesehurdlesstructurallynotjustviafinancial

supportiscritical,andthecostofdelayingactioncompoundseveryyearthatthegapwidens.

Thelong-termgrowthoutlookforAIanditssupportinginfrastructureisunambiguouslystrong,andthe

countriesandcompaniesthatpositionthemselves

decisivelywillshapetheeconomicsoftheAIera.Withthatframinginmind,thisstudyrevisitsfourquestionsthatstakeholdersacrossthedatacenterecosystemconsistentlyraiseonbuild-outmomentum,

ecosystemrisks,thepublicpolicyimperative,anddivergingregionaltrajectories.

2|Theglobalbuild-outrace:AIdatacenterstudy|Part1

Contents

P4

P12

P18

P25

P31

CoverAIGENERATED

P32

1/TheAIinfrastructureboom

1.1/Twoforcesdrivingdatacenterdemand

1.2/Scale,trajectory,andthecapitalunderwritingthebuild-out

1.3/ThephysicalrealityofAI-scaleinfrastructure

2/Structuralboom,cyclicalfragility-Potentialcorrectionrisksahead

2.1/Thefinancinggapbeneaththegrowthstory

2.2/Mixedmarketsignals,supplychainrisk,andthecaseforstructurallong-termgrowth

3/USdominance-WhyEuropeisfallingbehindintherace

3.1/USdominanceandthescaleofthegap

3.2/StructuralconstraintslimitingEuropeanbuild-out

4/ThepublicpolicyimperativeforEurope

4.1/ThepolicyrationalefordomesticAIinfrastructure

4.2/Frompolicyambitiontostructuralreform

Conclusion

Keyecosystemstakeholderquestions

Theglobalbuild-outrace:AIdatacenterstudy|Part1|3

TheAI

infrastructureboom

Theglobalbuild-outrace:AIdatacenterstudy|Part1|5

1.1/Twoforcesdrivingdatacenterdemand

F

ewsectorshavemovedfromspecialistinfrastructuretomainstreamconversationasrapidlyasthedatacenterindustry.Whetherit'sthescrambleforpowercapacity,theextraordinaryfinancialperformanceof

ATrainingcompute(FLOP)overtime

enablingcompaniessuchasNVIDIA,growingcommunityoppositiontonewdevelopments,orevenproposalsfororbitaldatacentersinspace,theconversationstartswiththeacceleratingadoptionofAItechnologies.

1e26

1e24

1e22

1e20

1e18

1e16

1e14

1e12

Gr

ok-4

FLAN137B

GPT-

DBR

4o

X

Al

phaGoZero

Meena

UL2

RD

T-1B

AlphaGoFan

GNMT

Libratu

ESM1

XLNet

Ga

s

Octo-

b

●

Base

S

MLM

VGG19

BIDAF

DeiTB

MMLSTM

S4

DITTO

AlexNetRNNLM

4.4x/year

GANs

RNTN

P

luribus

Swift

ILM●

NPLM

(Brown)

LMICA

●

GPUDBN

s

VImagegeneration

DRHNISS

Invariant

●

SparsevisCNN

ionencoding

Deeplearning

era

2000200320062009201220152018202120242025

Publicationdate

Source:EpochAI,secondaryresearch

TheAIrevolutionhasdrivendatacentergrowthintwokeyways.Firstisthecompoundingcomplexityofmodeltraining.Withtheentryintothedeeplearningerain2010,theamountofcomputedeployedinmodeltraininggrewbyafactoroffouryearoveryear.Thistrendcontinuedandacceleratedwiththearrivaloftransformer-basedlargelanguagemodels(LLMs).So-called"scalinglaws"describehowmodelperformanceimprovesasthenumberofparameters,thevolumeoftrainingdata,andthenumberoftrainingiterationsincreases.TheprogressionacrossGPTgenerationsservesasanexample:GPT-1launchedwith110millionparameters;GPT-4reportedlyoperatesatanestimated1.7trillion;GPT-5projectionsrangefromthreetofivetrillion.Weexpectsuchscalingdynamicstocontinuethroughatleast2030,asimprovingtheintelligenceandcapabilitiesofLLMsrequiresexponentiallymorecomputeresources.SatyaNadella,CEOofMicrosoft,describedthelinkbetweenmodelcapabilityandexponentialneedforcomputeresourcesinaninterview:"Ifintelligenceislogofcompute,whoevercandolotsofcomputeisabigwinner."A

TheseconddriveristhepaceandvarietyofAIadoption.TheblastofAIadoptioncan,inhindsight,betracedtotheyears2022and2023.Sincethen,demandforAIhasnotgrownincrementally;ithasoutpacedvirtuallyeveryprevioustechnologyshift.ChatGPTreached100millionmonthlyactiveusersinapproximatelytwomonths–fasterthanTikTok(ninemonths)orInstagram(twoandahalfyears)–addingbetweenoneandtwomillionusersperdayintheprocess.Thattrajectorydidnotreflectpent-updemandforafamiliarproduct;itwasthecreationofanentirelynewcategory.Byearly2026,weeklyactiveusershadreached883million,havingsurgedroughly70%injustfivemonthsduring2025beforestabilizing.Whatwouldtypicallyrepresentadecadeofgrowthwascompressedintoamatterofquarters.B

ThebreadthofAIadoptionmaymatterevenmorethanitspace.Enterpriseusecasesremainlargelyinearlydeployment,compute-intensiveworkloads–reasoning,agenticAI,andvideogeneration–arescalingrapidly,andtheindustryisshiftingtowardinferencecategoriesthataresignificantlymoredemandingthanthetext-basedapplicationsthatdrovetheinitialwave.

BAveragenewusersaddedperdayonthewayto100millionusers[m]

1.6

1.2

0.8

0.4

0.0

ChatGPTcameoutofnowhereand

addedaround1–2millionusersperday

onitswayto100million,compressing

decadesofadoptionintoweeks

Eachdigitalwavehasarrived

fasterthanthelast:

ChatGPThit100millionusersin

ca.2months,iPhonetookca.3

years,Facebookc.10years

ComputersInternetFacebookiPhoneChatGPT

Source:TheConferenceBoard

6|Theglobalbuild-outrace:AIdatacenterstudy|Part1

Theglobalbuild-outrace:AIdatacenterstudy|Part1|7

1.2/Scale,trajectory,andthecapitalunderwritingthebuild-out

T

hedatacentersector'scurrentgrowthphasemarksadramaticstepchange–acceleratingtheriseoflargeITcampusesalreadyreshapingtheindustry.Between2015and2020,theindustryunderwentasignificantexpansiondrivenbytheshifttocloudcomputingandtheproliferationofvirtualized,cloud-nativeusecases–theperiodthatestablishedtheinitialscaleofhyperscalerssuchasGoogleandAmazonWebServices.Thatfoundationmatters,butAI-specificusecasesarethedominantgrowthdriverstoday,

andareexpectedtoremainsoforatleastthenextdecade.Datacentercapacityistypicallymeasuredingigawatts(GW)ofpowerconsumption,reflectingtheenergyandcoolingneedsofthedatacenterinfrastructure,ratherthanitscomputecapacity.Thisstudyusestheso-called"nameplatepowerconsumption"andencompasseson-premiseenterprisedatacenters,colocationfacilities,andhyperscalercampuses.C

CProprietaryRBmodelforeseesgrowthto~220GWin2030and~340GWin2035

Evolutionofnameplatefacilitydatacentercapacity[inGW]

317

341

294

247

271

203

201

187

224

169

152

118

134

113

172

144

99

102

117

78

94

58

86

113

35

78

88

137

131

124

19

69

107

70

67

64

61

57

20252026202720282029203020312032203320342035

CAGR'30-'35

CAGR'25-'30

Total21%9%

44%

12%

AI

Non-AI10%5%

Evolutionofnet-addcapacityandreplacement[inGW]

Total18%6%

54

23

48

50

23

44

49

55

24

37

24

34

23

10%

~0%

39

24

30

29

26

29

22

Net-add

24

26

13

14

9

41

32

17

Replacement26%10%

20252026202720282029203020312032203320342035

Source:RolandBergeranalysis

8|Theglobalbuild-outrace:AIdatacenterstudy|Part1

TheRolandBergerDataCenterModelprojectsanexpansionfrom88GW–withAIrepresentingonlyapproximately20%ofthatload–in2025,to224GWwitha50%AIshareby2030,and341GWwith60%AIshareby2035.Themodelisbasedonchipproductionavailabilityforecasts,revenueexpectationsforkeychipmakerssuchasNVIDIAandBroadcom,aswellashyperscalerCapExestimates.Acrossarangeofpublishedforecasts,consensusisbroadlyalignedthrough2030forAI-linkedpowerdemandgrowingata35-45%compoundannualgrowthrate(CAGR),drivingtheAIshareoftotaldatacenterload

toroughly50%.TheRolandBergermodelextendsthisviewto2035,projectinganoverallcompoundannualgrowthrate(CAGR)of21%from2025to2030,and9%from2030to2035.Thesefiguresarealsoconsistentwiththeresultsofoursurveyofseniordecision-makersspanningEurope'sdatacentervaluechain,whichindicatedgrowthexpectationsof16%to22%between2025and2030and~10%for2030-2035.

Thecapitalfuelingtoday'sgrowthismainlyprovidedbyasmallnumberofhyperscalers.CombinedCapExacross

DSelectedhyperscaler1CapExevolution

TotalCapEx,2020-2028[USDbn]Forecast

1,000

800

600

400

200

0

937

884

755

412

239

158154

131

97

202020212022202320242025·202620272028

AmazonAlphabetMicrosoftMetaOracle

1SelectedhyperscalersshownincludeAmazon,Google,Microsoft,Meta,andOracle;broaderdatacenter&AIinfrastructureestimatesmaydifferdependingoninclusionofcolocation,neocloud,andpowerinfrastructurespend

Source:S&PCapItalIQPro

Theglobalbuild-outrace:AIdatacenterstudy|Part1|9

Amazon,Alphabet,Microsoft,Meta,andOracle–relativelyflatthrough2022–jumpedsharplyasAIbecamethedominantinvestmentthesis.Theindividualcommitmentsspeakforthemselves.AmazonexpectstoinvestaroundUSD200billioninCapEx,weightedtowardAWSforAIandcoreworkloads.GooglehasannouncedUSD180–190billionin2026alone.MicrosoftplannedtoincreaseitstotalAIcapacitybyover80%in2025,anddoubleitsdatacenterfootprintoverthefollowingtwoyears.However,newplayers,includingAIlabs,neocloudoperators,andcolocationproviders,arenowenteringtheCapExcycle,

broadeningthebaseofcapitalflowingintothesector.

D,E

ThegrowthiscascadingthroughsemiconductorcompaniessuchasNVIDIAandTSMC,whichhaverecordedsubstantialrevenuegrowthfrom2024to2025–aswellasdatacenterinfrastructuresuppliers.Eaton'smanagementdescribedabacklogequivalenttotwelveyearsofwhatwasbuiltin2025,withthepipelinecontinuingtogrowfasterthantheindustrycanbuild.SiemensreporteddatacenterrevenueincreasesinQ1offiscalyear2025/26ofaround35%.

EEvolutionof2026CapExestimatesthroughout2025and2026

Evolutionof2026forecast,01/2025–07/2026[USDbn]

1,000

800

600

400

200

0

738737755

667674

554

494

422429

300326337342352

440

509

528

291

361

01/2503/2505/2507/2509/2511/2501/2603/2605/2607/26

AmazonAlphabetMicrosoftMetaOracle

Source:S&PCapItalIQPro

10|Theglobalbuild-outrace:AIdatacenterstudy|Part1

1.3/ThephysicalrealityofAI-scaleinfrastructure

M

odernAIfacilitiesaredefined,aboveallelse,bycomputeandpowerdensity.Overthepastdecade,theindustryhasundergoneastructuralshiftinhowmuchcomputepowerandenergyconsumptionispackedintoagivenphysicalfootprint–andthatshifthascascadingconsequencesforeverylayeroftheinfrastructure.Rack-levelpowerhasmovedfrom10kilowatts(kW)inlegacyinstallationstoover100kWtoday,withfutureprojectionsreaching1megawatt(MW),whilefacility-leveldemandhasscaledfromafewmegawattstocampusesnowplannedinthegigawattrange.

Thisdensificationisnotjustafunctionofthedemandformorecomputepowerandtheeconomiesofscalewithinadatacenter,itisakeytechnicalrequirement.Efficientlargemodeltrainingandhigh-speedinferencebothdependontightlycolocatedprocessing,withminimallatencyandcommunicationoverheadbetweenunits.Infact,today'srackscontainingcloseto100graphicsprocessingunits(GPUs)aredesignedtofunctionasasinglecomputeunitforthealgorithmstheyrun.

ThedensityofmodernAIcomputeisdemandingonitssupportinginfrastructure.Theindustryhasshiftedfromairtomoreeffectivebutmorecomplexliquidcoolingastheprimarymethodofmanagingrack-levelheat–achangethataffectsfacilitydesign,floorloading,andplumbinginfrastructurefromthegroundup.Further,innovationsindirect-currentpowerdistributionwithindatacenters,whichreduceconversionlossesandcopperneedsandsimplifydeliverytohigh-densityracks,aremovingfromdevelopmentintopracticaldeployment.

Theimplicationsofthisarchitectureonenergyinfrastructurearesubstantial.Aonegigawattdatacenterdrawspowerequivalenttotheaverageconsumptionofacityofonemillionhouseholdsor,arguingfromthegenerationside,absorbsthepoweroutputofaconventionalWestern-stylenuclearpowerplant.

Securingthatpowerinvolvesthreedistinctbutrelatedchallenges:

?Themereavailabilityofpowerfromtheelectricitynetworkoron-sitepowergenerationfacilities

?The"right"gridaccess(increasinglytothehigh-voltagetransmissiongrid)andon-sitepower

distributioninfrastructurerequiredtodeliveritreliablytothecomputeandnon-compute(e.g.,coolingandventilation)equipment

?Thecapabilityofthegridandon-sitefacilitiestodealwithpotentiallyhugelyoscillatingpowerdemand

withinthedatacenter.

Trainingrunsfrequentlyextendbeyond100consecutivedays,imposingextremelyhighavailabilityandreliabilityrequirementsonthepowerinfrastructure.Compoundingthis,datacentersdonotdrawpowerataconstantrate.Asafacilitycyclesbetweencompute-intensiveandcommunication-intensivephasesduringatrainingrun,itsaggregatepowerdemandcanoscillatemateriallywithinseconds–placingextremestressonboththegridconnectionandon-siteinlineorstandbybackupsystemsdesignedtoabsorborbridgethosefluctuations.Aselectricitycostsaccountfor~40%ofdatacenteroperatingexpenditure(OpEx),dependingonthelocation,thecombinedchallengeofcostsandgridaccesshaspromptedseveraldatacenteroperatorstoinvestinon-sitegenerationcapacity.

ThesupplychainservingthegrowingAIdemandisunderconstantstressfrommultiplefronts.GPUshortagesremainaliveconstraint,withMetasecuringtensofbillionsinNVIDIAcommitmentspreciselybecauseproductioncapacitycontinuestostrainagainstdemand.Gridaccessisincreasinglyconstrained–forexample,intheUKthere

Theglobalbuild-outrace:AIdatacenterstudy|Part1|11

isasignificantshortfallofpoweravailabletosupportplanneddatacenters,andintheUnitedStates,gasturbinemanufacturingleadtimeshaveextendedtofiveyears,ashyperscaler-ledpowerprocurementreshapesglobalequipmentsupplypriorities.Thesearenottemporaryhurdles;theyarestructuralfeaturesofanindustrybuildingatapacethatphysicalinfrastructurecycleswerenotdesignedtoaccommodate.

HyperscalerCapEx–some

USD750billionin2026,mostlyfordatacenters–isstrainingtheglobalsupplychain,fromtransformers,turbines,and

coolingtochips.●●

Datacenter101

Questionsmostpeoplewanttoaskbutneverdo

IfAIlivesinsoftware,whydoeseverybodytalksomuchaboutphysicaldatacenters?

Today'sgenerativeAIisbuiltonlargelanguagemodelswithhundredsofbillions–sometimestrillions–ofparameters.TheseparametersaretheAI.Producingthemrequiresaprocesscalledtraining:runningvastdatasets throughthemodelrepeatedlyovermorethan100consecutivedays,ina single,integratedfacility.Oncetrained,modelsmustalsohandleinference–respondingtouserqueriesatscale–whichexceedstheprocessingpower,memory,andenergycapacityofindividualdevicesandbenefitsenormously fromeconomiesofscale.ManyAIlabsalsochoosetokeeptheirmodelscentralizedratherthandistributethemtoendusers,reinforcingtheneed forlarge,purpose-builtinfrastructure.

Structuralboom,

cyclicalfragility-

Potentialcorrectionrisksahead

Theglobalbuild-outrace:AIdatacenterstudy|Part1|13

2.1/Thefinancinggapbeneaththegrowthstory

T

hecapitalandrentalcommitmentsunderwritingAI'sinfrastructurebuild-outarerunningwellaheadoftherevenuesandcashflowsavailabletosupportthem.DespiterevenueofUSD13billionin2025,OpenAIreportedlygeneratedacashburnofUSD8billion.Theshortfallwascoveredbyfurtherexternalfunding.Similarly,AnthropiccommittedUSD50billioninCapExattheendof2025againstUSD9billioninrevenuerunrateatthistime,requiringexternalfundingliketheUSD30billionseriesGinFebruary2026.Bothcompaniesare,ineffect,financingtheirinfrastructureambitionsandexpensesontheexpectationoffuturerevenuegrowthatarateandscalethathasnodirectprecedentinprivatemarkets.

Thisisnot,initself,asignofanindustryindistress.Land-grabphasesinhigh-growthtechnologysectorsroutinelyinvolvespendingaheadofrevenue,andbothcompaniesareoperatinginanenvironmentwheretechnologicalleadershipcompoundsquickly.Theriskisnotthegap–itisthesensitivityofthemodeltoitsownassumptions.AsDarioAmodei,CEOofAnthropic,notedinaFebruary2026interview:"IfI'mjustoffbyayearinthatrateofgrowth,orifthegrowthrateis5xayearinsteadof10xayear,thenyougobankrupt."

Hyperscalersareinastructurallystrongerposition,fundingAICapExpredominantlyfromthecashflowsgeneratedbytheirlegacycloudbusinesses.Butevenhere,themathistightening.Freecashflowforseveralofthemajoroperatorsisexpectedtoturnnegativein2026,andsuccessiveroundsofexpandingCapExguidancehavebeenmetwithgrowingcautionfrominvestors.ThemarkethasnotwithdrawnitsconvictioninAI,butitisbeginningtoaskwhentheinvestmentphaseconvertstoreturns,andonwhattimeline.F

Hyperscalers'relianceondebtmarketsisgrowing.In2025,hyperscalersissuedmorethanUSD120billionindebt–againstanannualaveragebelowUSD30billionperyearbetween2020and2024.AmazonaloneplacedUSD

54billionindebtinMarch2026.Thisisnotinherentlyproblematicforcompaniesofthisscaleandcreditquality,butitmarksameaningfulstructuralshiftinhowthebuild-outisbeingfinanced.Equallynotable,theaggregateCapExexpectationfor2026morethandoubled–fromapproximatelyUSD350billiontoUSD750billion–betweenJuly2025andJuly2026.Ifthatpatternofupwardrevisioncontinues,theassumptionsunderpinningcurrentreturn-on-investmenttimelineswillrequirerevisiting.

AICapExisracingaheadof●revenue.Cashflowfrom

othercorebusinessearnings

andexternalfinancingbridge

thegapfornow,butmomentumwillholdonlyifbreakevenof

revenuesandcostscomeswithinreach.

FHyperscalerCapExandoperatingcashflow(OCF),2020-2029

Actualsandconsensusestimates[USDbn,%]

Amazon

2023202420252026202720282029

CapEx/OCF62%72%94%108%92%75%65%

301

387

250

224

221

240

140132

186200

8553

11683

Alphabet

OCFCAGR30%CapExCAGR56%

CapEx/OCF32%42%56%88%91%77%59%

317

253

384

211

231

242

187

165

125

102

91

53

32

228

OCFCAGR

CapExCAGR

Microsoft

%%

CapEx/OCF34%44%52%79%77%72%63%

355

229

225

199

176

161

158

126

282203

83

56

35

103

OCFCAGR

CapExCAGR

Meta

CapEx/OCF38%41%60%98%95%80%70%

211

168160

137133

116

91

168263183

71

70

27

37

OCFCAGR

CapExCAGR

Oracle

CapEx/OCF41%53%259%215%196%169%124%

96

7390

77

49

36

2235

2011

177

5999

1

OCFCAGR

CapExCAGR

Sum1

CapEx/OCF41%50%68%98%94%80%67%

1463

977

940884

769755

1168937

478

412

377

603

154

239

%%

OCFCAGR

CapExCAGR

Source:Companyfilings,S&PCapItalIQPro1SumnotdrawntoscaletoenhancelayoutOCFCapEx

14|Theglobalbuild-outrace:AIdatacenterstudy|Part1

Theglobalbuild-outrace:AIdatacenterstudy|Part1|15

Datacenter101

Questionsmostpeoplewanttoaskbutneverdo

WhocapturesthemostupsidefromtheAIdatacenterexpansion?

AIchipcompanies–ledbyNVIDIAandBroadcom–capturethelargestshareofnear-termvalue,asGPUsandAIacceleratorssitatthecoreofeverydatacenterandcommandpremiumpricingandmargins.Memorymanufacturersareclosebehind:AIchipsrequirehigh-bandwidthmemory(HBM)tomovedatarapidlyduringtrainingandinference,andsupplyremainstightenoughtoconstraintheentirebuild-out.Thesecondpartofthisstudyexamineswhereelseinthesupplychainsignificantvalueisbeingcreated.HyperscalerssuchasAmazon,Microsoft,andGooglebuythishardware,deployitatscale,andrentAIcapacitytodownstreamclients–includingmodeldeveloperslikeOpenAIandAnthropic–whointurnusethatcapacitytobuildproductsandservices,thoughmanyhaveyettoreachsustainableeconomics.

2.2/Mixedmarketsignals,

supplychainrisk,andthe

caseforstructurallong-termgrowth

P

otentialmarginpressurecompoundsthefinancingrisk.On-demandandspotGPUrentalrateshavebeenfluctuating,makingtheeconomicsofthepureGPU-as-a-Servicemodeldifficulttopredictforinfrastructureplayers("neoclouds")thatdependonthissegment.Atthesespotpricinglevels,themodelispotentiallyathin-marginbusiness,requiringtightcontroloverenergycosts,CapExplanning,andhardwarelifecyclemanagement.ThisappliestoindividualGPU-hourspotandshort-termon-demandrates.Forlargerinstallations–particularlytrainingclustersandincreasinglyinferencefleets–thepicturelooksdifferent:scarcityandlong-termcommitmentslockingincapacityyearsinadvancewereevidentinspring2026,asdemonstratedbytheAnthropic,xAI,andAWSagreements.Therationale:growingdemandforever-largertraininginstallationsandAnthropic'ssurgefollowingitsClaudeCodesuccess.

Thesupplychainstructureitselfintroducesafurtherlayerofinstability.Duetothelengthofprocurementcyclesandthetimerequiredtobuilddatacenters,modestshiftsinenddemandcanproducedisproportionatelylargeswingsinupstreamordersfordatacenterequipmentsuchasGPUs,memory,andcoolingequipment.Eachst