为AI供电：随着AI派对继续，自带电源（BYOP）成为数据中心洞察v3.0

Powering AI: Bring Your OwnPower ('BYOP') as AI PartyContinues - Data Center Insightsv3.0 Thematic Investing William Thompson+1 212 526 8641william.s.thompson@barclays.comBCI, US AI infrastructure remained a key theme across Tech,Industrial and Energy earnings. With speed-to-power key inthe AI arms race, emphasis turns to bring-your-own-power(BYOP) solutions as the energy happy hour faces last call. Weagain compile recent Powering AI insights with 50+ companymentions. Hannah Greenberg+1 212 526 2457hannah.greenberg@barclays.comBCI, US Hiral Patel+44 (0)20 3134 1618hiral.patel@barclays.comBarclays, UK Laia Marin i Sola, CFA+ 44 (0)20 3134 2531laia.marin@barclays.comBarclays, UK The AI infrastructure party continues...Hyperscalers again raised AI infrastructure capex asdata center demand continues to outpace supply. We highlight 1) hyperscaler capex trends, 2)strong public cloud demand, 3) elevated data center pre-leasing, 4) proliferation of AI trainingclusters, 5) surging AI inference demand, 6) US data center construction spend, and 7)hyperscaler electricity use. Katherine Ogundiya+44 (0)20 3134 1391katherine.a.ogundiya@barclays.comBarclays, UK ...but speed-to-power fuels bring-your-own-power ('BYOP') solutions.Data center energydemand continues to outpace the ability to add grid power and infrastructure. This is leading toon-site power policies, including in Texas and Ireland, as speed-to-power fuels demand fordistribution power across turbines, engines, fuel cells and energy storage, which we summarizein Figure 16. We compile companies exposed to the distributed power theme in Figure 17. Amy Lian+44 (0)20 3134 0179amy.lian@barclays.comBarclays, UK Company mentions include:Tech (AAPL, AMZN, CRWV, GOOG, META, MSFT, ORCL); Datacenters(CORZ, DLR, EQIX, IRM); Industrials (ABB, AAON, CARR, CAT, CLF, CMI, ENR, ETN, FIX,GEV, LR, MOD, NVT, OC, RR, SIE, SU, TSLA, TT, URI, WRT1V); and Energy (AEP, BE, BKR, CPK,GNRC, SEI, WMB). Siddharth Gupta+91 22 6175 2540siddharth.gupta1@barclays.comBarclays, UK April 2025 survey of data center leadersSource: Bloom Energy The AI Infrastructure Party Continues.... . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4Hyperscaler Capex Continues to Climb. . . . . . . . . . . . . . . . . . . . . . . . . . . . .4Public Cloud Underpins Data Center Demand. . . . . . . . . . . . . . . . . . . . . . . 8Strong Pre-leasing Supports Colocation Expansion. . . . . . . . . . . . . . . . . . 9Proliferation of Large AI Training Clusters. . . . . . . . . . . . . . . . . . . . . . . . . . 11Surging AI Inference Demand. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12US Data Center Construction Spending. . . . . . . . . . . . . . . . . . . . . . . . . . . . 13Hyperscaler Electricity Use. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 ...But the Energy Happy Hour Faces Last Call. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17Pivoting to Bring-your-own-power ('BYOP') Solutions. . . . . . . . . . . . . . . 17Distributed Power was a Major Earnings Theme. . . . . . . . . . . . . . . . . . . . 24 Related Research. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 The AI Infrastructure Party Continues... We outlined our rather bullish AI infrastructure views in Powering AI Chart Pack Summary (28Slides) and Deep impact: The global AI arms race, which included why we thought AI was beingunderhyped (not overhyped), why the paradigmshiftin AI scaling laws was leading toincreasing AI computing and energy intensity despite hardware and algorithmicefficiencies(e.g. DeepSeek [private, not covered]), and the critical role of AI infrastructure and energy in theglobal AI arms race to general intelligence (AGI) and artificial superhuman intelligence (ASI). The post-DeepSeek bear thesis for AI infrastructure that emerged earlier this year suggestedthat ongoing AIefficiencieswill continue to bend the AI compute and cost curves andcommoditize AI, and this would lead to 1) hyperscalers pulling back on AI-directed capex amidlow AI returns; 2) downward revisions to energy growth forecasts; and 3) an oversupply of datacenters, as we detailed in Powering AI: Post-DeepSeek Bull/Bear Debate. This belief was alsoreinforced by speculation that MSFT was slowing and pausing leasing activity, which proved tobe more noise than anything, as we note in MSFT data center plans. However, hyperscalers haveraised (not cut) AI infrastructure-related capex and signaled out-year capex growth based on thefollowing factors: 1.Continued data center supply/demand tightness with power being the primaryconstraint.Companies emphasized that strong cloud and AI demand continues to outpacesupply, with Amazon (AMZN, covered by Ross Sandler) noting that power is the single-largest supply constraint. 2.Optimisms on AI returns.Companies are already seeing strong, measurable