For three years the binding constraint on AI was accelerators. That constraint has not disappeared, but it now shares the critical path with something less portable: electricity, and the ability to connect a large load to a grid in a reasonable time.
What changed
Training clusters moved from megawatts to hundreds of megawatts, and inference adds a persistent load that never idles. A single campus can now consume what a mid-sized city draws. Grid interconnection queues in several markets stretch to multiple years, which makes power availability a siting decision rather than an engineering afterthought.
The three responses
- Contract for generation: dedicated supply, including nuclear agreements and long-term renewables deals.
- Move to where power is cheap: new campuses where generation is stranded and land is available.
- Change the workload: better utilisation, lower-precision inference, and scheduling jobs against power price and temperature.
Why efficiency does not solve it
Efficiency per token keeps improving, and total demand keeps rising faster. This is the classic rebound pattern: cheaper inference expands the set of things worth doing. A model that costs a tenth as much does not reduce consumption by ninety percent; it creates nine new use cases.
The local politics
Data centre approvals now run into local resistance over water use, noise and grid strain. The projects that proceed tend to be the ones that bring their own generation, use closed-loop cooling, and can show a tax base. Utilities, meanwhile, are being asked to fund upgrades for a single customer, which is a regulatory argument that will run for years.
What it means for everyone else
Industrial electricity prices, grid reliability, and the pace of electrification all interact with this demand. The AI build-out is now a power-sector story as much as a software one, and the countries that can add generation quickly will host the next round of capacity.
Bottom line: the next bottleneck is not a chip shortage. It is a connection queue, a cooling loop and a planning committee, and those move on a slower clock than any model release.
