SemiAnalysis: 40GW+ of behind-the-meter US datacenters by 2028

A quiet shift in how the US builds hyperscale AI datacenters is now visible in the numbers, and the way SemiAnalysis lays it out, the country's grid isn't the primary bottleneck for AI capacity anymore. Behind-the-meter power is. The headline projection is 40GW+ of behind-the-meter (BTM) datacenter capacity in the US by 2028, powering more than half of new datacenters that year, with total BTM equipment TAM crossing 50GW per year by 2029.

The grid math is what forces the shift. PJM's 2027/2028 base residual auction cleared 134,478 MW of unforced capacity at a 14.4% reserve margin against a 20% target, and the report describes grid headroom as turning negative by 2027. Gas nameplate additions run under 10GW per year in 2026-2027 and only pick up in 2028+, while gas turbine and generator step-up transformer lead times have blown out to 3-4 years versus a historical 18 months. Combined-cycle plants sit on a 4-6 year development timeline. Hyperscalers are going behind the meter because the front of the meter can't move fast enough for a 2027-2028 in-service date.

ERCOT is where the co-location activity is most concrete: roughly 2,885 MW of announced projects, including Crusoe's Goodnight Campus at 525.5 MW co-located with 933 MW of gross gas generation permitting, AWS at Comanche Peak with a 1,200 MW co-located load on a 20-year PPA ramping to full capacity by 2032, CyrusOne's Thad Hill at 400 MW across two phases, and a CyrusOne/Constellation Freestone project up to 760 MW. Meta's Ohio campus is reportedly designed never to connect to the PJM grid, and its self-built AI sites target roughly two nines of uptime rather than the traditional five, trading reliability for speed.

The honest caveat is that BTM speed hinges on the same turbine and transformer queues that are already strained. SemiAnalysis notes PJM's queue saw 24GW of projects with fully executed agreements terminate since 2020, and 29% of milestone changes between January 2023 and January 2026 came from permitting. What the reporting doesn't give you is the environmental permitting exposure for dedicated gas builds, or how single-customer plants get financed if AI capex softens before contracts mature. NERC's 2025 assessment already flags 13 of 23 North American areas for resource-adequacy shortfalls over the next decade, so whoever moves fastest behind the meter today may still be exposed to grid stress tomorrow.

The upside sits with gas turbine OEMs, transformer suppliers, and integrated developers like Crusoe. The more interesting change is that hyperscalers with land, gas access, and balance sheets big enough to post multi-billion-dollar performance letters of credit now have a speed moat that regulated utilities can't match. Watch whether Meta's design-never-to-connect stance spreads from Ohio to the rest of the AI buildout.