Power Is the New Bottleneck for AI

In 2023 the scarce input for AI was the accelerator. Waiting lists for GPUs stretched for quarters and the industry's attention fixed on foundries and packaging capacity. That constraint has eased. The one replacing it is harder to manufacture: electricity, delivered where it is needed, at the scale and reliability a gigawatt-class campus demands.

The International Energy Agency has documented data centres as a fast-growing component of electricity demand, concentrated in a handful of regions. In those regions, the queue to connect a large new load to the grid is measured in years, transformer and switchgear lead times have lengthened, and utilities are being asked to plan for a customer class whose load profile they have never seen.

Why now

• The scale of AI campuses has grown from tens of megawatts to plans sized like power plants.
• Inference demand is becoming continuous as agents run around the clock.
• Interconnection processes were designed for generation projects, not sudden large loads.
• Technology companies with strong balance sheets can now sign decades-long power agreements.
• Ratepayer advocates and regulators are asking who bears the cost of grid expansion.

What changes

Data centre developers are turning into energy developers. Site selection used to prioritise latency, fibre and tax incentives. It now starts with a question about substation capacity and ends with a power purchase agreement. Hyperscalers are contracting directly for nuclear output, backing new gas plants and funding grid upgrades themselves to skip the queue.

Behind-the-meter generation, in which a campus produces its own power and consumes it directly, is moving from exotic to expected. It shifts the risk profile of a project: the developer now carries fuel, permitting and emissions exposure that used to sit with the utility.

A second design pattern is flexible compute. Training runs and batch inference can, in principle, follow power availability across sites and hours. That makes AI a potential grid asset rather than only a burden, but it requires software and contracts that do not yet exist at scale.

Who benefits, who is at risk

Beneficiaries: manufacturers of transformers, switchgear and cooling; grid operators with spare capacity; nuclear and storage developers; regions with cheap, clean, firm power. At risk: developers in grid-constrained markets, residential ratepayers if upgrade costs are socialised, and corporate net-zero commitments that assumed flat electricity demand.

The numbers to watch

We deliberately avoid quoting single headline figures for AI electricity demand. Estimates vary widely by methodology and horizon, and the Parallax AI Infrastructure Index will publish its own tracked figures with sources. What is not in dispute is the direction: demand is rising faster than grid capacity in the regions where AI wants to be built.