Nvidia has come up with an unusual way to combat the AI energy crisis, and it demands even more GPUs again

The boom in AI has bumped up against an unexpected limiter: ordinary power delivery. Not a chip shortage this time, but substations and feeders that strain under the peak loads of massive training and inference farms.

NVIDIA’s response is a bit lateral. By year‑end it plans, with partners, a pilot of roughly 25 small data centers (DCs) sited close to electrical substations across the US. The concept: don’t chase huge, contiguous capacity — instead shift jobs around the grid. If one substation is pegged, workloads move to another that has slack.

What matters here are tiny pockets of spare power. Think free segments of 5–20 megawatts (5–20 MW), amounts that won’t run a hyperscale campus but will power a compact AI cluster, e.g., a few racks tuned for model training or inference bursts. The goal is operational flexibility rather than gross energy savings.

Mark Spieler, NVIDIA’s Senior Director of Energy, points out there are about 55,000 substations in the US; even modest reserves at many nodes add up into a nontrivial compute pool. That arithmetic is hard to ignore.

There’s an ironic twist: decentralizing like this increases hardware needs. You need reserve capacity, duplicate GPUs, orchestration to flip workloads fast — so the fix shifts the problem toward inventory, ops complexity and latency tradeoffs. I can’t help feeling it’s a neat workaround that buys capacity at the cost of a bigger logistical headache.