Bernstein: Only 34% of North America's 586GW AI Data Center Pipeline Is Credible
nashnova research
Bernstein's tracker puts North America's data-center pipeline at 586 GW, yet only 199 GW — roughly 34% — passes its credibility screen. This means about two-thirds of announced capacity cannot yet be counted as reliable future supply.
586 GW sounds enormous — how much will actually get built?
Bernstein's Nov 10 report shows the pipeline swelled from 411 GW in July to 586 GW in October — a 43% jump in three months.
Risk-adjusted credible capacity barely moved: 136 GW (33%) → 170 GW (35%) → 199 GW (34%).
This means → the headline number keeps inflating, but the credible share stays locked at one-third. Two-thirds of planned capacity remains on paper.
How different are "announced" projects from "under construction"?
About 497 GW (85% of the pipeline) is still at the "announced" stage; probability-weighted credible capacity is just 123 GW, a 25% credibility rate.
Projects already under construction total only 87 GW, yet their credible capacity reaches 75 GW — an 86% credibility rate.
In plain terms = the best predictor of whether a data center will actually exist is not the press release — it is whether dirt is moving on-site.
Which developers have the most credible projects?
Hyperscale cloud companies (Amazon, Google, etc.) building their own facilities score highest at about 75% credibility; specialist colocation operators reach roughly 62%.
Emerging AI-cloud providers sit at only 29%; crypto-mining pivots at 24%; other developer types average just 11%.
This reflects a clear pattern: turning a PowerPoint into a live data hall still requires deep capital, firm customer commitments, and proven construction experience.
How dominant are the Big Four tech companies?
Amazon, Google, Microsoft, and Meta together account for about 78 GW of credible capacity — 39% of the credible total.
By company: Amazon leads at ~33 GW, Google ~20 GW, Meta ~15 GW, Microsoft ~10 GW.
This means → despite a wave of newcomers announcing multibillion-dollar AI data-center plans, nearly four-tenths of credible capacity sits with just four companies. Industry concentration is far higher than the headlines suggest.
Will tighter regulation kill these projects?
Texas — the largest U.S. data-center market — has about 150 GW in its pipeline, but Bernstein sees only 43–44 GW as credible.
Roughly 21 GW in Texas is already under construction and largely insulated from new restrictions; pre-construction projects face potential grid-access and permitting delays.
In plain terms = regulation is more likely to stretch timelines and raise costs than to cancel projects outright.
What should investors watch?
Bernstein's top picks are Equinix (EQIX) and Digital Realty (DLR), both rated outperform with price targets of $1,270 and $226 respectively.
On the equipment side, the firm favors Quanta Services (PWR), Eaton (ETN), and Hubbell (HUBB).
This means → the persistent two-thirds gap between announced and credible capacity is the variable that will test whether the AI-infrastructure investment thesis actually delivers — investors should track who is building, not who is talking.
市场有风险,内容仅供研究参考,不构成投资建议。
