AI's Real Bottleneck Isn't Chips. It's Electricity.

Every conversation about AI's future seems to be about chips. Who has the most GPUs, whose model is biggest, which lab is compute-rich. But spend an hour reading energy data instead of model benchmarks and a different bottleneck shows up — one you can't fix with a better architecture or a bigger order from a chip foundry.

AI doesn't just need compute. It needs power. An astonishing, grid-bending amount of it. And the grid, unlike a data center, can't be spun up in a quarter.

The numbers are genuinely startling

According to the IEA, global data centers used about 415 terawatt-hours of electricity in 2024 — roughly 1.5% of all electricity on Earth. That already sounds like a lot. Then you look at the trajectory.

One estimate has data-center consumption approaching 1,050 TWh as soon as 2026. To put that in perspective: if data centers were a country, they'd be around the fifth-largest electricity consumer in the world — sitting between Japan and Russia. The IEA's base case sees them hitting 945 TWh by 2030 and 1,200 TWh by 2035, with electricity demand from AI-optimized data centers more than quadrupling by 2030.

We've been measuring the AI race in parameters and FLOPs. The honest unit might be terawatt-hours.

The hard part isn't generating power — it's delivering it

Here's the bit that reframed it for me. The problem isn't only "make more electricity." It's getting that electricity to a specific patch of land at a specific time.

The IEA points out that in some legacy data-center hubs, there's simply no spare capacity left on the local grid — leading to official and de-facto moratoriums on new builds. Meanwhile the unglamorous physical guts of the grid — transformers, switchgear — now have lead times that have roughly doubled. You can have the GPUs on a truck and still wait years for the hardware that connects them to power.

That's a very different kind of constraint than the ones AI people are used to. You can't optimize a transformer backlog with a clever kernel.

Why this changes how I think about the industry

A few consequences I keep landing on:

1. Efficiency stops being a nice-to-have. When power is the ceiling, a model that does the same work for half the energy isn't just cheaper — it's more deployable. The "small model that's good enough" suddenly has a structural advantage.
2. Location becomes strategy. Compute migrates to where power is — next to hydro, nuclear, stranded renewables. The map of AI starts to look like the map of the energy grid.
3. AI and energy policy are now the same conversation. You can't plan one without the other, even though they live in completely different worlds today.

The closing thought

The romantic version of AI progress is smarter models. The real version, increasingly, is substations, transmission lines, and who can get a gigawatt connected before their competitor. It's the least sci-fi story imaginable, and it might be the one that actually decides how far this goes.

The bottleneck was never going to be imagination. Turns out it might just be the wire.