US utilities are demanding proof that AI data centres are real

US utilities are asking AI data centre developers to prove that their proposed projects are real before committing scarce grid capacity. The shift follows a surge of connection requests that vastly exceeds the amount of demand utilities expect to materialise, according to reporting summarised by TechRadar Pro. ¹

The phantom data centre problem

US developers have reportedly requested 1,066 gigawatts of grid connections, but Wood Mackenzie expects utilities to commit to only about 28% of that demand. Some of the gap comes from developers filing the same project with several utilities while they compare delivery times. Other applications may come from speculators holding a queue position without firm site control or a signed customer.

That uncertainty matters because grid upgrades are not free. New substations, transmission lines and other equipment can be planned around an expected customer, while the cost is ultimately shared across the wider network if that customer never arrives. Utilities are therefore asking for collateral, longer contracts, security agreements and engineering study fees before they agree to build around a proposed data centre.

Exelon shows what stricter checks reveal

Exelon offers one example of the effect. After the utility began requiring transmission security agreements and related credit obligations, its high-probability data centre load fell from 18 gigawatts at the end of 2025 to roughly 11GW in the second quarter, while its wider pipeline under study dropped from about 43GW to 25GW. Exelon has said its $41 billion capital plan through 2029 remains unchanged, so the reduction in expected load has not yet translated into an equivalent cut to planned spending.

The pressure is particularly visible in PJM, the largest US grid operator. Its capacity auction has reportedly cleared at the price cap three times in a row while missing its reliability target by 6.8GW. A market monitor also told federal regulators that 4,017MW of forecast data centre load had already been cancelled or delayed. ¹

That creates a difficult balancing act. Utilities cannot wait forever for perfect certainty when transformers and other grid equipment can take years to arrive. Yet building too aggressively around speculative demand risks leaving households and existing businesses to share the bill. Average minimum contracts for large customers have reportedly risen to around 10.8 years, up from 4.6 years before 2024.

Our opinion

Utilities are right to ask harder questions before turning speculative AI demand into permanent infrastructure. A queue position is not the same thing as a working data centre, and a spreadsheet full of optimistic forecasts should not automatically become a bill for households. The danger is that the industry swings too far in the other direction: genuine projects could be delayed because the process becomes so cautious that nobody can secure power in time.

The sensible answer is not to reject AI infrastructure or accept every application. It is to make commitments progressively: require meaningful financial guarantees at each construction milestone, publish clearer queue data, and make developers pay for the risk created by projects that disappear. That would give utilities a firmer basis for planning without pretending that every proposed gigawatt is real.

Hero image: server racks photographed by panumas nikhomkhai on Pexels. ²