Power is the new bottleneck
Elon Musk warned that electricity supply is becoming the constraining factor for AI buildouts, with chip production outpacing global power‑supply growth and causing multi‑year waits for grid connections at datacentres. If power availability and permitting become the limiting factor, scaling training capacity will depend as much on siting and utility agreements as on silicon production. (x.com) (x.com)
The old artificial intelligence bottleneck was chips. The new one is the thing behind the wall that keeps the chips on: electricity. (iea.org) Elon Musk has been warning that power, not silicon, is becoming the limiting factor for new artificial intelligence clusters, and regulators are now rewriting grid rules around “large loads” above 20 megawatts because data centers have become too big for old connection processes. (ferc.gov) A data center is just a warehouse full of computers, but the newest artificial intelligence warehouses draw power like heavy industry. The International Energy Agency says a typical artificial intelligence-focused data center uses as much electricity as 100,000 households, and the largest sites under construction will use 20 times that. (iea.org) The reason is simple: each new rack of chips is turning into a denser electric heater that also happens to do math. Hewlett Packard Enterprise says one NVIDIA GB200 NVL72 rack consumes 132 kilowatts, with most of that heat requiring liquid cooling instead of the air cooling older server rooms used. (buy.hpe.com) Chip companies can ship that hardware faster than utilities can deliver new substations, transformers, and transmission lines. A July 2024 U.S. Department of Energy advisory group said hyperscale facilities asking for 300 to 1,000 megawatts of service were already stretching local grids, with lead times of 1 to 3 years just to supply power. (energy.gov) And that is before the wider grid catches up. Lawrence Berkeley National Laboratory says the median time from a generation project’s grid connection request to operation reached 5 years for projects built in 2023, which means even new power plants are slow to arrive when data centers need them. (proceedings.bpa.gov) The scale of the demand jump is now showing up in national forecasts. The International Energy Agency projects global data center electricity use will rise from 415 terawatt-hours in 2024 to about 945 terawatt-hours in 2030, which is slightly more than Japan’s entire electricity consumption today. (iea.org) In the United States, the Electric Power Research Institute says data centers could reach up to 9% of total electricity generation by 2030, up from roughly 4% today. That is why utility planning dockets now treat server farms less like office parks and more like aluminum smelters. (prnewswire.com) Virginia shows what this looks like on the ground. Dominion Energy’s latest filings say data center-related power requests in Virginia have reached 70,000 megawatts, which Virginia Business reported is about triple Dominion’s current peak load. (virginiabusiness.com) Texas is seeing the same scramble from a different angle. The Electric Reliability Council of Texas said it created a dedicated “large load integration” process, and officials said in late 2025 that more than 70% of the state’s 233 gigawatts of large-load requests were coming from data centers. (ercot.com) (utilitydive.com) So the race is changing shape. Winning no longer just means getting more graphics processing units; it means finding a county with land, a utility with spare capacity, a regulator willing to approve upgrades, and a power contract that arrives before the hardware is obsolete. (energy.gov) (ferc.gov)
