Power for AI: which supply options can arrive before 2030
Simon Irish, CEO of small modular reactor developer Terrestrial Energy, published an opinion piece in Utility Dive on Nov. 18, 2026, arguing that the grid will fall short of AI's power needs and that no single resource fixes it.
The numbers he cites are stark. McKinsey estimates the U.S. will need 30 GW to 55 GW of additional capacity by 2030, driven by data centers and aging plant retirements. PJM Interconnection's latest capacity auction came in about 6.6 GW below its reliability requirement and cleared at the price cap. On the supply side, GE Vernova's gas turbine backlog and slot reservations reached 116 GW at midyear 2026, and the company says a heavy-duty turbine ordered now will not arrive until 2031. On-site generation at data centers may avoid some grid upgrades but does not speed up equipment deliveries, so it faces the same shortage.
Some options move sooner. A DOE initiative targets 5 GW added at existing nuclear plants by 2029, and a $1.9 billion DOE loan would restart Iowa's Duane Arnold plant, returning 615 MW. Duke University's Nicholas Institute estimates the grid could absorb 76 GW of new load if that load accepted curtailment averaging 0.25% of annual consumption, in events of about 1.7 hours. Wind, solar and storage deploy quickly, but round-the-clock supply needs substantial overbuilding, storage and transmission.
For abundance, the point is that cheap compute depends on cheap, available power, and constrained supply raises its price. Irish recommends matching solutions to their timelines, offering earlier interconnection to customers who accept enforceable curtailment, and signing reservation or offtake agreements now for long-lead firm clean power. He says advanced nuclear will not close the gap this decade; its test is cost at fleet scale. DOE modeling points to 550 to 770 GW of additional firm clean capacity by 2050, with about 200 GW from advanced nuclear.
Caveats matter. This is an opinion piece by an executive with a commercial stake in nuclear. He concedes forecasts can be wrong and requested load can be speculative or duplicative, that many data centers have shown little willingness to curtail AI workloads, and that few operating nuclear plants remain candidates for significant uprates, with restarts underway at nearly every practical site.
Source: Utility Dive
Many Minded