AI's real ceiling is megawatts
Blackwell racks at 120–140 kW. Gas proposals for US data centers hit ~189 GW. Moody’s sees halls at ~10% of US power by 2030. Firm watts — not GPU rumor — set the pace.
By Drew Wall,
The model race still sounds like chips. The buildout is watts. Hopper racks sat near 40 kW; Blackwell NVL72 is 120–140 kW; Rubin quotes climb past 190 kW. Air cannot cool that. Moody's now puts US data-center load near 426 TWh by 2030 — about 10% of national electricity, up from ~5% today. Agents and bigger training runs only raise the draw.
Density, then the queue
Liquid cooling is a first-order design requirement — loops, CDUs, leak detection — not a retrofit. You can order GPUs faster than you can hook a 100–500 MW campus to the grid. Interconnection waits still run years. Speed to power beat latency and fiber as the site filter.
Firm watts — mostly gas for now
Renewables still get bought. Firm capacity gets bought harder. Global Energy Monitor put ~189 GW of gas proposed next to US data centers by mid-2026 — roughly double six months earlier — with Texas alone near 77 GW. Behind-the-meter turbines and engines are the near-term crutch while nuclear PPAs (TMI restarts, Susquehanna, SMR offtakes) stay on multi-year clocks. Heat rejection and water stay on the same spreadsheet as kilowatts.
The point
The real limit on AI data centers is electricity, not chips, so plan and invest as if each one were a power plant that happens to run models. For the spending side, see profitability. For local projects being blocked, see cancellations. Related categories: Energy, Cloud GPU, Infrastructure.