AI Data Centres on Track to Use Two-Thirds of US Households' Electricity by 2030 — and PJM Is Already Six Gigawatts Short

The numbers have moved out of the abstract. By 2030, AI data centres alone will consume the equivalent of the electricity currently used by two-thirds of all US households, according to industry forecasts compiled in early May 2026. Total data-centre electricity consumption is projected to double by 2030; AI-specific consumption is set to triple. The existing US grid is not built for it.

The PJM warning

PJM Interconnection is the largest US grid operator, serving more than 65 million people across 13 states from New Jersey to North Carolina. PJM forecasts a 6 GW shortfall against its reliability requirements by 2027 — a gap large enough that the grid's independent market monitor has called the situation "crisis stage right now." PJM has never been this short.

How operators are responding

Three patterns. First, behind-the-meter generation: hyperscalers and AI-native operators are co-locating power generation with compute, increasingly with natural gas turbines, nuclear restarts (Three Mile Island, Palisades) and a swelling pipeline of small modular reactor offtake agreements. Second, dedicated transmission: utility-scale projects with private financing for the transmission lines connecting compute to generation, often crossing state regulatory boundaries that make this harder than the technical problem. Third, demand-side management: data centres bidding into wholesale markets and curtailing load during peak hours, a posture that has gone from theoretical to economically necessary.

The nuclear pipeline

The most striking number: small modular reactor offtake agreements between data-centre operators and SMR projects have grown from 25 GW at the end of 2024 to 45 GW by early May 2026. Most of those agreements are conditional and most of those reactors will not produce power before the early 2030s. But the pipeline is now structurally different from anything the US nuclear industry has had since the 1980s. Four executive orders signed earlier this year by the White House to accelerate nuclear deployment have added political tailwind.

The politics

Polymarket traders put a 93.5% implied probability on at least one qualifying AI data-centre moratorium passing into US state law by year-end 2026. Local opposition has consolidated around water consumption, grid stress, ratepayer impact (data-centre load can push residential bills higher when interconnect costs are socialised) and emissions. Several states are debating moratoria; some county-level moratoria are already in place.

What it means in Europe

Europe is on a parallel but later trajectory. EU data-centre demand is rising, particularly in Frankfurt, Dublin, Amsterdam and increasingly the Nordics. Luxembourg sits in a useful position — a stable grid, cool climate corridors near the Moselle, fibre density — but grid capacity is the binding constraint here as it is in the US. The IEA flagged this in its 2025 review of data-centre electricity surge. Decisions taken in 2026 about generation, transmission and zoning will define what the European AI compute footprint looks like in 2030.