版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领
文档简介
July2025
Mckunsey
&company
Whatisadatacenter?
AdatacenterisafacilitythathousesandrunsITinfrastructurethat’scriticaltothedigitaleconomy,particularlygenAI.
WhenyoutypeaquestionintoagenAIplatform,youreceiveananswersofastthatitmay
feellikemagic.Ofcourse,it’snot:The
modelsthatmanyofusrelyon
,bothpersonallyand
professionally,aretheresultofdecadesofresearchandtrillionsofdollarsofinvestments—nottomentionvastandever-increasingamountsofenergy.
DatacentersarespecializedfacilitiesthatmanageITinfrastructure,includingservers,storagedevices,andnetworkequipment.Theyplayacriticalroleinprocessing,storing,anddistributinglargeamountsofdata,makingthemessentialto
genAI
andtherestofthedigitaleconomy.
McKinseyanalysisindicatesthatby2030,datacenterswillneed
$6.7trillionofworldwide
investmenttokeeppacewiththedemandforcomputepower—around70percentofwhichwillcomefromAIworkloads(Exhibit1).“Overthenextdecade,”saysMcKinseySeniorPartner
PankajSachdeva
,“theindustrywillgothroughan
S-curveofdemandgrowth
tosupportthe
infrastructurethatwillpowerthedigitalrevolutionandcontinuetopowerthe
cloudrevolution
.”
Notonlywillexistingdatacentersneedtobecomemorepowerful,butnewdatacenterswillalsoneedtobebuiltapace.Howcantheworldmeetthisquicklygrowingdemand?FindoutthisandmoreinthisMcKinseyExplainer.
LearnmoreaboutMcKinsey,s
Technology,Media&TelecommunicationsPractice
,
Whatiscomputepower?
Computepowerisemergingasoneofthisdecade’s
mostcriticalresources
.TheriseofAIhasledtoskyrocketingdemandforcomputepower,orthehardware,processors,memory,storage,andenergyneededtooperatedatacenters.
Whatarethecorecomponentsofadatacenter?
Therearefourcorecomponentstomostdatacenters:
—ITequipment,Datacentershostservers,storagedevices,andnetworkdevicesthathandledataprocessing,storage,andtransmissionsneeds.
—Infrastructureandutilities,Datacentersareequippedwithair-conditioning,redundant
electricitysystems,andelectricityconditioningtoensureuninterruptedoperations.
—Connectivity,Datacentersaretypicallylocatednearhigh-bandwidthfibernetworksthatenablelow-cost,high-speeddataexchange.
—Physicalsecurity,Robustphysicalsecuritymeasures,includingfiresuppressionsystems
andrestrictedaccess,aretypicallyimplementedtoprotectthecenter’sequipmentanddata.
LearnmoreaboutMcKinsey,s
Technology,Media&TelecommunicationsPractice
,
Whatisadatacenter?2
Whatisadatacenter?3
Exhibit1
BothAlandnon-AWorkloadswillbekeydriversofglobaldatacentercapacitydemandgrowththrough2030.
Estimatedglobaldatacentercapacitydemand,'continuedmomentum'scenario,gigawatt
219
3.5x
2025-30change
Non-AIWorkload
AIWorkload
2025
2026
2027
2028
2029
2030
2025-30total
IncrementalAlcapacityaddedperyear,gigawatts
13
21
22
31124
Note:Figuresmaynotsumtototals,becauseofrounding.
source:MckinseyDatacenterDemandModel;Gartnerreports;IDCreports;Nvidiacapitalmarketsreports
Mckinsey&company
HowisAIinfluencingthegrowthofdatacenters?
AI’sboomhasfueled
skyrocketingdemandforpower
.Asnotedabove,70percentoftheprojecteddemandfordatacentercapacitywillcomefromAI-basedworkloadsby2030.
McKinseyanalysissuggeststhatinamidrangescenario,demandforAI-readydatacenter
capacitywillriseatanaveragerateof33percentperyearbetween2023and2030.GenAI,
currentlythefastest-growingadvanced-AIusecase,willaccountforaround40percentofthetotaldemand(Exhibit2).
TokeepupwiththerapidriseofAI,
datacentershavebecomebiggerandmorepowerful
.Tenyearsago,acenterwith30-megawattcapacitywasconsideredlarge;today,a200-megawattcampusisconsiderednormal.AI-readydatacentersconsumeanespeciallylargeamountof
Exhibit2
Whatisadatacenter?4
Alisthekeydriverofgrowthindemandfordatacentercapacity·
EstimatedglobaldatacentercapacityDemandforadvanced-Alcapacity,'
demand,'gigawatts%oftotaldatacentercapacitydemand
250
200
150
50
20232030
20232030
1MidrangescenarioisbasedonanalysisofAadoptiontrends;growthinshipmentsofdifferenttypesofchips(application-specificintegratedcircuits,graphics
processingunits,etc)andassociatedpowerconsumption;andthetypicalcompute,storage,andnetworkneedsofAIWorkloads.Demandismeasuredbypowerconsumptiontoreflectthenumberofserversafacilitycanhouse.
source:MckinseyDatacenterDemandmodel
Mckinsey&company
energybecauseoftheirhighaveragepowerdensities—thatis,theenergyconsumptionofserversintheracks.Averagepowerdensitieshavemorethandoubledinjusttwoyearsandareexpectedtorisenearlyfourtimesby2027.
HyperscalersincludingAmazonWebServices,GoogleCloud,MicrosoftAzure,andMetaarethecompaniesfuelingmostoftoday’sincrementaldemandforAI-readydatacenters.That’sbecausethesehyperscalersrequiremassivecapacitytohostboththelargemodelsthey
developin-house,suchasGoogle’sGemini,andthosedevelopedbyAIcompanies,suchas
OpenAI’sChatGPT.Cloudserviceproviderscurrently
ownmorethanhalf
theworld’sAI-readydatacenters.McKinseyestimatesthatby2030,upto65percentofAIworkloadsinEuropeandtheUnitedStateswillbehostedonhyperscalers’infrastructure.
Mostothercompaniesareusingoff-the-shelfmodelsthatarelargelyhostedonapubliccloud.
ButasAImatures,moreorganizationsarelikelytobuildandtraintheirownmodelsbasedoninternaldata,whichcouldincreasedemandforprivatehosting.
Whatisadatacenter?5
WhatdonewAI-readydatacentersrequire?
Thehigherenergydemandandpowerdensity,aswellasthecomplexityofdifferentAIworkloads,areleadingtorapidchangein
threemainareasofdatacenterconstruction
:
—Locationandpowerinfrastructure,Asdatacentersproliferate,powersupplyisbecominganissueinmarketsthathavetraditionallyattractedclustersofdatacenters,suchasNorthernVirginiaandSantaClara,California,intheUnitedStates.Manyutilitiesfindthattheyhaven’tbeenabletobuildtransmissioninfrastructurequicklyenough,raisingconcernsthatthey
maybeunabletogeneratesufficientpowerinthefuture.InadequatepowergenerationcanslowdatacenterexpansionandaffecttheoverallconsumerandbusinessuseofAI.
—Mechanical(cooling)systemdesign,AIserversconsumesomuchenergythattheyget
physicallyhot,somuchsothatair-basedcoolingsystemscan’tkeepup.Thisissuehas
promptedashifttoapproachesthatremoveheatdirectlyfromracksbyusingliquid,whichismoreefficientatabsorbingandtransferringheatthanair:forexample,rear-doorheat
exchangers,direct-to-chiptechnology,andliquidimmersioncooling.
—Electricalsystemdesign,AIworkloadscallforlargerpowerdistributionunitsthatcancopewithhigherpowerdensities—andinresponse,manydatacenteroperatorsareinstalling
largerswitchgearandfloor-mounteddistributionunits.Thesechangesreducethe
complexityaswellasthecapitalandoperationalcostsofinstallingandmaintainingmultiplesmallerunits.
LearnmoreaboutMcKinsey,s
Technology,Media&TelecommunicationsPractice
,
Dodatacentersalsosupportnon-AItasks?
AIworkloadsmaydominatetheconversation,butnon-AIprocessingloadsandcloudmakeupa
significantportionofdatacenteractivity
.TheseworkloadsincludetraditionalenterpriseITtaskssuchaswebhosting,enterpriseresourceplanningsystems,email,andfilestorage.Non-AItasksrequirelesscomputepowerandcanoperateefficientlyoncentralprocessingunits,ratherthanthespecializedgraphicsprocessingunitsorAIacceleratorsthatAIworkloadsrequire.
Non-AIloadsalsotendtohavemorepredictableusagepatternsandlowerpowerdensities,
whichallowforlessdemandingcoolingandenergyrequirements.Asaresult,datacentersthatfocusonnon-AIprocessingtypicallyhavedifferentinfrastructureneeds,capitalintensity,andoperationalconsiderationscomparedwiththosethatareintendedprimarilyforAIworkloads.
Whatisadatacenter?6
Whatregionalchallengesdoesthedatacentersectorface?
InEurope,thedatacentersectorfaces
severalchallenges
,includinglimitedsourcesofreliablepower,sustainabilityconcerns,insufficientpowerinfrastructure,landavailabilityissues,
shortagesofpowerequipment,andalackofskilledelectricaltradespeople.Inmajormarkets,itcantakeuptofiveyearsormoretosupplypowertonewdatacenters,andthepowergridis
increasinglystrained.Meetingthesedemandswillrequirealotofcleanenergy,whichwillinturnrequiretheconstructionofmorenewenergysystemsthatcanbeturnedonintimesof
especiallyhighdemand.
IntheUnitedStates,thedatacentersectorfacesmoresignificantchallenges,particularlyintermsof
powerconnectionsandlaborconstraints
.Thereisalsoashortageofelectricaltradeworkers,whichaffectstheabilitytoexecuteprojectsontimeandcandelaythebuildingof
datacentersandassociatedpowerinfrastructure.Tariffshavealsoincreasedanelementofuncertaintyandcouldpresentadditional
supplychaincomplexities
.
LearnmoreaboutMcKinsey,s
Technology,Media&TelecommunicationsPractice
,
Howcanenergyplayersgetinvolvedinthedatacentersector?
Investorshave
plentyofopportunities
toparticipateandcanenablesolutionsforpoweraccessandsources.Hereare
fourhigh-potentialareas
:
—Secondarymarketswithaccesstoreliable,cheappower,There’satiminggapbetween
datacenterbuilds—whichcanbedonein18to24months—andpowerinfrastructure
development,whichcantakeanywherefromthreetotenyears(sometimesevenmore)to
complete.Buttherearecreativewaystobridgethisgap.Manyhyperscalersarebuildingoutcapacityinnew,nontraditionallocationsoutsidecoredatacentermarketsbecausethese
areashaveaccesstocheaper,availablepoweraswellasthepotentialtobuildcarbon-freeinfrastructure.IntheUnitedStates,Iowa,Wyoming,Indiana,andOhioeachhouseorhavereceivedinvestmentsfromatleasttwoofthetopfourhyperscalers.
—Behind-the-metersolutions,Thesesolutionsprovidepowerinareaswhereutilities
providerscannotkeepupwiththepaceofdemandorreliabilityrequirementsas
transmissionconstraintsortheavailabilityoflocalpowersupplyworsen.Forexample,
thereareopportunitiesforinvestorstobuildpowerthatcanbefullyislandedoutsidethegridorprovidesupplementalpower(suchasnuclear)tocomplementtheexistinggrid.
Whatisadatacenter?7
—Sustainabilityambitionsdrivenbyrenewable-energyproviders,Hyperscalersthathave
madeclimatecommitmentswillrequirehundredsofterawatt-hoursofcleanenergytomeetfuturedemand.Solarand
onshorewind
areexpectedtogeneratemostoftheworld’snewcleanenergy,butothercleanenergytechnologiescanalsosupplyenergyinthemediumto
longterm.Thesesourcesinclude
offshorewind
,nuclear,
geothermal
,gas,carboncapture
and
storage
,andcleanfuels.
—Transmissionanddistributioninvestments,Poweravailabilityiscriticallyimportantto
meetingdatacenterdemand,andutilitycompaniesaretakingnotice.Investorscanfunnel
investmentsintoutilitycompaniestobuildouttransmissionanddistributioninfrastructureinkeymarkets.
LearnmoreaboutMcKinsey,s
Technology,Media&TelecommunicationsPractice
,
Whatrolecanrealestateorganizationsplayinthedatacenterindustry?
Datacenterspresent
threemainopportunities
torealestateorganizations.Atthemostbasiclevel,realestateorganizationscanbuythelandfordatacenters—specifically,parcelsofland
thattheythinkwillgetthepower,networkconnections,andcustomerdemandneededfordatacenters—thensellorleasethislandtodatacenterdevelopers.
Realestatecompaniescanalsodopartialdatacenterdevelopment,thenselltheirfacilitiesto
developersorfinalcustomers.Finally,companiescancreatea“co-location”model:Theybuytheland,constructthebuilding,provideaccesstopowerandconnectivity,andbuildouttheinterior.Thesefacilitiesareoftenleasedbyenterprisesorhyperscalers.
“Comparedtootherrealestateassetclasses,”saysMcKinseySeniorPartner
PankajSachdeva
,“datacentershistoricallyhavehadhigheryield.Duetosupply-sideconstraints,wedon’t
anticipatethatrealestateplayersordatacenterdeveloperswillfacemajoryieldcompressionsoverthelongterm.Realestateinvestorsareseekingexposuretodatacenterstogethigher
growthand
sustainedlevelsofhigheryield
.”
HowcantelecomoperatorscapitalizeonthegrowingdemandforAIinfrastructure?
Beyondprovidingtheinfrastructurethatpowerscommunicationandconnectspeople,telecomoperatorscanbuildthe
infrastructurethatwillrealizeAI’sfullpotential
.Onewayislayingdownfastinternetcables(orfibertoconnectnewdatacenters,whichwillhelppeopleandbusinesses
Whatisadatacenter?8
access
powerfulcloudservices
).Telecomoperatorscanalsoturnunusedspaceintoprofitby
offeringAIcomputingpower,knownasgraphicsprocessingunits,orGPUs,asaserviceforrent.Anotheroptionisbuildingandrunningtheirowndatacenterstosupportthehighcomputing
powerandinternetspeedthatAIrequires.Theycouldalsoaddressadditionalgapstobridgeconnectivityplatforms.
Becausetheyalreadyhavewidecoverageandexperiencemanagingnetworks,telecom
companiesareinastrongposition.Buttheycanalsopartnerwithcompaniesthatmanufacturecomputerchips,builddatacenters,orofferothertechservices.
It’simportanttohaveaclearaspirationandambitionforcapabilitiesandinvestmentneededtosucceed,especiallyasoperatorsgofurtherupthestackfromthecoreconnectivity
infrastructure.
Toreducerisk,telecomfirmscanstartsmallbyinvestinggraduallyinAI-readyinfrastructure
andmakinguseofwhattheyalreadyhave.Clearcommunicationandanimbleapproachwillbekeytosuccess.
LearnmoreaboutMcKinsey,s
Technology,Media&TelecommunicationsPractice
,Andcheckout
jobopportunitiesrelatedtodatacenters
ifyou,reinterestedinworkingwithMcKinsey,
Articlesreferenced:
—“
Thecostofcompute:A$7trillionracetoscaledatacenters
,”April28,2025,
Jesse
Noffsinger
,
MarkPatel
,and
PankajSachdeva
,withArjitaBhan,HaleyChang,and
Maria
Goodpaster
—“
AIinfrastructure:Anewgrowthavenuefortelcooperators
,”February28,2025,
AbhyuadayaShrivastava
,
BrendanGaffey
,
GustavGrundin
,
SebastianCubela
,and
Tomás
Lajous
,withAkshatAgarwal,DapoOrimoloye,LorraineSalazar,MiguelFrade,andNicholasShaw
—“
Howhyperscalersarefuelingtheracefor24/7cleanpower
,”December18,2024,
Lorenzo
MoaveroMilanesi
,
TjarkFreundt
,and
YuitoYamada
,withFridolinPflugmannandMarc
Ludwig
—“
AIpower:Expandingdatacentercapacitytomeetgrowingdemand
,”October29,2024,BhargsSrivathsan,
MarcSorel
,andPankajSachdeva,withArjitaBhan,Haripr
温馨提示
- 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
- 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
- 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
- 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
- 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
- 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
- 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
最新文档
- 中药材种植员安全检查测试考核试卷含答案
- 康复辅助技术咨询师创新应用评优考核试卷含答案
- 经编工道德评优考核试卷含答案
- 压榨机工工作效率知识考核试卷含答案
- 水泥混凝土制品养护工岗中生产安全培训考核试卷含答案
- 用电检查员安全素养评优考核试卷含答案
- 印染成品定等工沟通技巧强化考核试卷含答案
- 灯具零部件制造工岗前安全防护考核试卷含答案
- 模锻工改进能力考核试卷含答案
- 压缩天然气场站运行工岗前总结考核试卷含答案
- 2026年秋季开学新教师校园安全责任入职培训
- 医院迁建项目高压开关柜安装施工方案
- 中石油俄语水平考试试题及答案
- 2026年浙江省综合性评标专家库评标专家考试在线题库
- 替奈普酶治疗急性缺血性卒中共识2026
- 2025年广东佛山仲裁委员会选聘仲裁员笔试备考题库附答案详解
- 医院后勤服务外包管理规范
- 重庆市2026年普通高等学校招生全国统一考试调研(四)语文试卷
- 广西壮族自治区公共资源交易平台系统权益类交易子系统(土地使用权)交易中心用户手册
- 光明网社招笔试题
- 上海市莘庄中学等四校联考2026届数学高一下期末经典试题含解析
评论
0/150
提交评论