Powering AI strains U.S. grid
- Wood Mackenzie says AI data-center buildouts are turning electrical gear into the bottleneck, with U.S. equipment demand rising from $20 billion to $65 billion by 2030. - The pinch point is basic hardware, not exotic chips: annual transformer demand could jump from about 1,500 units now to more than 9,000. - Grid stress is already showing up in power markets, with PJM auctions and utility forecasts repricing how fast AI campuses can actually open.
The constraint on AI is starting to look a lot less like chips and a lot more like electricity. Not just generation, either — the boring middle of the system matters now: transformers, switchgear, substations, feeders, interconnection studies. That is the new bottleneck. The immediate news is that Wood Mackenzie now expects the U.S. data-center electrical equipment market to more than triple by 2030 because hyperscale AI buildouts are colliding with a grid supply chain that was never sized for this. (datacenterknowledge.com) ### What exactly is getting squeezed? The physical gear that takes bulk power and makes it usable inside a campus is getting scarce. Wood Mackenzie pegs the market for U.S. data-center electrical equipment at roughly $20 billion today and about $65 billion by 2030, with transformer demand alone rising from around 1,500 units a year to more th(datacenterknowledge.com)ans a project can be financed and half-built but still sit there waiting for metal in a yard. (datacenterknowledge.com) ### Why is AI worse than old-school cloud? AI facilities are denser and hungrier. Wood Mackenzie says U.S. data-center capacity could jump from roughly 24 GW to 100 GW between 2026 and 2030, and rack densities are pushing toward 200 kW. Higher density changes everything downstream — electrical distribution, backup design, cooling, and how muc(datacenterknowledge.com)loads that hit the grid like a new factory town. (datacenterknowledge.com) ### How big could this get? EPRI’s latest scenarios say data centers use about 4% to 5% of U.S. electricity now, but could reach 9% to 17% by 2030 and 10% to 20% by 2035. Virginia already stands out — data centers there consume more than 25% of electricity today, and EPRI’s scenarios put that share at roughly 39% to 57% by 2030. Several other(datacenterknowledge.com). (powering-intelligence.epri.com) ### Where is the stress showing up first? In queues and prices. PJM’s capacity market exists to secure enough future supply for reliability, and the recent auctions show how tight things are getting. The 2027-28 auction cleared at the $333.44 per MW-day cap after PJM came up about 6,625 MW short of its reserve-margin target. PJM also said a 5,250 MW increase in its demand forec(powering-intelligence.epri.com)blem. But it does mean new load is now big enough to move regional power economics. (pjm.com) ### Why does the “boring hardware” matter so much? Because generation without delivery is useless. A gas plant, solar farm, or small reactor does not help a data center if the substation is late, the transformer is missing, or the interconnection upgrade is stuck. Wood Mackenzie says data centers could account for as much as 40% of total U.S. electrical-equipment demand by 2030. Utilities, factories, ren(pjm.com) The biggest buyers can lock up supply early. Smaller buyers and even utilities get pushed back in line. (datacenterknowledge.com) ### So what does this change for data-center design? It pushes operators toward power-aware architecture. If power delivery is uncertain, campuses need staged energization, regional failover, and degraded modes that keep priority workloads alive while shedding lower-value jobs. Admission control stops a cluster from promising compute it may n(datacenterknowledge.com)me when energization slips or utility limits tighten. This is an inference from the supply and grid constraints, not a direct forecast. (datacenterknowledge.com) ### Does this help gas and nuclear? In the near term, yes, mostly because they are dispatchable and familiar to grid planners. PJM’s cleared capacity mix in the latest auction was led by gas at 43%, followed by nuclear at 21% and coal at 20%, while very little new capacity showed up. That weak new-build response is part of why the market keep(datacenterknowledge.com)est rises whenever the grid starts saying “not fast enough.” (utilitydive.com) ### Bottom line? AI is no longer just a semiconductor race. It is a transformer race, a substation race, and a queue-management race. The winners will not just be the companies with the best models — they will be the ones that can actually get power to the rack.
