The world's power grids are being redesigned around artificial intelligence
Datacentre expansion has turned electricity supply into the binding constraint on the AI industry — and into a political problem for governments.
The short version
AI compute demand is growing faster than grid capacity can be added, pushing utilities, regulators and technology firms into a fight over who pays for new power infrastructure.
For most of the last two decades, the limiting factor in computing was silicon. It is increasingly electricity.
Datacentre operators building capacity for large AI models are signing power agreements that rival the consumption of mid-sized cities. That has changed the calculation for utilities, which historically planned for flat or slowly declining demand in wealthy economies.
Why the constraint moved
Training and serving large models concentrates load in a way that traditional industrial demand does not. A single campus can request connection capacity that would previously have been spread across an entire region, and it can request it on a timeline measured in months rather than the decade-long horizon grid operators are used to.
The result is a queue. In several markets, the wait for a new high-capacity grid connection is now longer than the wait for the chips that would fill the building.
The political dimension
When electricity becomes scarce, computing stops being a purely commercial question and becomes a public one.
Governments face a choice they have mostly avoided stating plainly: whether new generation capacity should be prioritised for industrial compute, for household supply, or for electrifying transport and heating. Each option carries a different price signal for consumers, and none of them is politically comfortable.
What to watch
- Interconnection reform, which determines how quickly new load can be attached to existing networks
- Behind-the-meter generation, where operators build their own supply and bypass the queue entirely
- Tariff design, which decides whether ordinary bill-payers subsidise industrial connections
- Siting decisions that follow cheap power rather than cheap land
The industry's public framing remains efficiency: better chips, better cooling, better scheduling. Those gains are real, and they are being consumed almost immediately by additional demand.
The bottom line
Compute is now an energy business. The companies that secure long-term, low-carbon, low-cost power will set the pace of the next model generation — and the regulators who decide how that power is allocated will have more influence over the technology's trajectory than any single laboratory.
Daily briefing