助力AI：我们追踪美国数据中心45+吉瓦的项目管道

Powering AI: We Track 45+ GW inthe US Data Center Pipeline The Powering AI theme is in vogue. The big question: what'sin the data center pipeline? We track 45+ GW of large USprojects over the next several years and outline 7 interestingobservations amid questions on capacity constraints. Thematic Investing William Thompson+1 212 526 8641william.s.thompson@barclays.comBCI, US Hannah Greenberg+1 212 526 2457hannah.greenberg@barclays.comBCI, US We track at least 45 GW of large US data center projects and it isn't an exhaustive list. Tracking DC projects and what is real vs. speculative is a full-time job. We've started with largeAI-focused US projects and track 45+ GW over the next several years, with the majority tied toStargate, hyperscalers, and xAi, though this isn't an exhaustive list and excludes many smaller/colo projects. Hiral Patel+44 (0)20 3134 1618hiral.patel@barclays.comBarclays, UK Hyperscalers continue to raise capex, but will it hit a power wall?More and more investorsseem to share our view that a grid power wall is forcing theshiftto distributed power and thatthisshiftisn't temporary. AI workload power fluctuations, politicization of higher utility bills,broader NIMBY and permit resistance, labor shortages, and equipment supply chains alsopresent risks to hyperscaler capex and AI ambitions. Laia Marin i Sola, CFA+ 44 (0)20 3134 2531laia.marin@barclays.comBarclays, UK Katherine Ogundiya+44 (0)20 3134 1391katherine.a.ogundiya@barclays.comBarclays, UK Amy Lian+44 (0)20 3134 0179amy.lian@barclays.comBarclays, UK Siddharth Gupta+91 22 6175 2540siddharth.gupta1@barclays.comBarclays, UK Executive Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3US Large Data Center Project Tracker. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 7 Interesting Observations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161. Stargate 1 Deploying Gas Despite Grid Capacity. . . . . . . . . . . . . . . . . . 162. Where are Crusoe's Other Turbines Going?. . . . . . . . . . . . . . . . . . . . . . . 163. All-the-above Approach to AI Power Fluctuations. . . . . . . . . . . . . . . . . 164. Beggars Can't Be Choosers on Power Equipment. . . . . . . . . . . . . . . . . 185. Opportunities to Skip the Turbine Order Queue. . . . . . . . . . . . . . . . . . 206. Emergence of Energy-as-a-Service (EaaS) Opportunities. . . . . . . . . . 227. Data Center Cost Data Points. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 Appendix. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23Related Reports. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 Executive Summary Track 45+ GW of large US data center projects and it's not an exhaustive list.With oursemiconductor analyst colleagues tracking over $3 trillion of announced compute spend(Scratch That, $3T and Counting – NVDA Pointing to Near-Term Numbers Upside), we'reoftenasked about the data center supply situation. Frankly, tracking projects and what is real vs.speculative is a full-time job, but we've started to compile large AI-focused US projects (at leastwhat has been publicly disclosed). We've started with large AI-focused US projects and track45+ GW over the next several years. This isn't an exhaustive list, nor does it include manysmaller data centers, including co-location facilities, which will add up. Furthermore, thenumbers largely reflect IT capacity—assuming a 1.2x Power UseEffectiveness(PUE) ratio,45+ GW of IT capacity would require 54+ of power capacity.We're happy to share our large USdata center project monitor file upon request. We can also share our list of ~150 companiestied to the Powering AI theme. We track 30+ GW tied to Stargate, the hyperscalers and Elon Musk's xAI (Fig 2, Fig 3, Fig 4).This includes 7 GW of Stargate capacity to date, though OpenAI (private, not covered) seems toexpect another 3 GW to join the Stargate family by the end of this year; Meta (META, covered byRoss Sandler) has announced up to 7 GW (though more likely 4 GW by 2030) across 3 AI-focusedprojects; Amazon (AMZN, covered by Ross Sandler) added 3.8+ GW in the past 12 monthsglobally ("more than any other cloud provider") with capacity up 2x vs. 2022 (25+% p.a.growth) and expects this to double again by 2027 (40+% p.a. growth ), which impliesanother ~18 GW by 2027, or ~13 GW assuming 70% in the US, based on our math (Figure4);Microsoft(MSFT, covered by Raimo Lenschow) is adding 900 MW in Wisconsin and has notedmultiple identical projects in other US locations; we assume there is another ~800 MW ofhyperscaler on-site power tied to William's (WMB, covered by Theresa Chen) $3.1bn investment;and xAI (private, not covered) is expanding to 1.4 GW in Memphis, Tennessee. We'll assume thatthis group accounts for at least 30 GW,