The US grid is growing fast. ICF International forecasts the country will add up to 445 gigawatts of generating capacity through 2030, a pace the consulting firm says could effectively add another grid as large as the country’s biggest regional power system. That’s a grid expansion story the energy sector has been reporting for months.
The deliverability story is different. ICF finds that only approximately 191 gigawatts of that capacity would be available as dispatchable peak-demand power, because intermittent wind and solar can’t be called on demand. That figure is roughly equivalent to the entire capacity of the PJM Interconnection, the largest US regional grid. ICF estimates the US has only 26 gigawatts of excess generating capacity above minimum reliability needs — roughly 3% of total US capacity. In the fastest-growing markets, Texas and PJM, ICF said there’s no spare capacity to support new demand beyond next year.
Warning
All three figures are sourced to ICF International's analysis as reported by EnergyNow. The article body was not directly readable in source verification. Treat these as reported findings pending confirmation against the full ICF publication.
That framework is directly relevant to AI infrastructure planning. Dispatchability — the ability to call on power when and where you need it — is what data centers require. A gigawatt of solar capacity that generates at peak midday production isn’t the same as a gigawatt of firm power available at 2 a.m. when a large language model is processing a batch job. Capacity numbers overstate what’s actually available for always-on, high-density compute loads.
The bottlenecks ICF identifies go beyond generation intermittency. Transmission deliverability constraints, including permitting delays and supply chain bottlenecks, could prevent much of the planned generation capacity from reaching AI data center demand centers. A new solar farm in the Southwest doesn’t help a data center in Virginia if the transmission interconnection queue is 18 months long and the substation upgrade is waiting on a transformer that’s backordered.
Why this matters for markets: Hyperscalers have been racing to lock in power capacity through long-term offtake agreements, direct grid connections, and on-site generation. The ICF analysis suggests the race isn’t just for capacity — it’s for dispatchable capacity, which is scarcer than the headline grid growth numbers imply. This is the third consecutive cycle where grid constraint reporting has featured in AI infrastructure news, following FERC’s show-cause actions against grid operators in the June 21 cycle. The pattern is consistent: power availability for AI data centers is tightening, not easing, even as the grid grows.
What to Watch
What to watch
the interconnection queue data. The Federal Energy Regulatory Commission publishes quarterly interconnection queue statistics. Watch whether the average time-to-interconnection for large commercial projects, which includes data center campuses, is increasing or holding steady through mid-2026. That’s the ground-level signal that confirms or contradicts the ICF deliverability warning. Physical AI infrastructure investment is accelerating even as the power constraint tightens. The companies that locked in firm power contracts 18 months ago are the ones with optionality right now.