US data center electricity demand is projected to jump 253% from 2026 levels to a record 194 gigawatts by 2035, The Kobeissi Letter reports, a figure roughly equal to the output of 194 traditional nuclear reactors. As a result, data centers are expected to account for about 20% of total US electricity consumption by 2035, up from roughly 6% today, with the share already projected to hit 12% by 2030. Much of the buildout is concentrated in a handful of grid regions, chief among them PJM, the interconnection that serves Washington DC and 13 states including Virginia, Pennsylvania and Ohio, where AI data center clusters are expanding fastest. The projection captures a shift already visible in utility earnings and grid planning: power availability, not chip supply, is increasingly the constraint that determines how fast hyperscalers can bring new AI capacity online. Utilities in PJM territory have flagged multiyear queues for new large-load interconnections, and grid operators warn that demand growth of this scale requires new generation and transmission capacity, and in some cases, direct on-site power deals that bypass the public grid entirely.
Why it matters · If electricity, not GPUs, becomes the hard ceiling on AI capacity growth, the winners shift from chipmakers to whoever controls power generation, transmission rights and grid interconnection queues.
Worth asking · Power grid capacity, not chip supply, may be the real bottleneck on the AI buildout through 2035. Does that make utilities and grid infrastructure the next AI trade, or is off-grid power the faster path?