The new bottleneck isn’t chips, it’s power

For the first two years of the AI infrastructure boom, the binding constraint was almost always GPU supply. Whoever could secure allocation, often years in advance, controlled the pace at which they could build. That constraint has not disappeared, but it has been joined, and in many cases overtaken, by a different one: power.

The numbers involved are now large enough to be genuinely difficult to reason about. A single gigawatt of AI data centre capacity is roughly equivalent to the output of a large power station, and individual operators are now talking about multi-gigawatt campuses as standard units of planning, not aspirational long-term goals. For context, a national grid the size of the UK’s peaks at somewhere around 35 to 40 gigawatts in total. Several individual AI infrastructure announcements over the past year have, on their own, represented a meaningful fraction of that.

This creates a planning problem that is fundamentally different from procuring GPUs. Grid connections take years to secure and build, are subject to planning permission and local consultation processes, and depend on transmission infrastructure that was not designed with this scale of point-load demand in mind. An organisation can, in principle, order GPUs and receive them within a procurement cycle measured in months. Securing the power to run them at scale operates on a completely different timeline, often the longer of the two constraints by a wide margin.

The industry’s response has been to attack the problem from several directions simultaneously. The most visible is behind-the-meter generation: building dedicated power generation, often gas turbines in the near term, directly alongside data centre campuses rather than relying solely on grid connections. This sidesteps grid queue delays but introduces its own planning, permitting and emissions considerations, and shifts the operator from being purely a compute business into also being, in effect, an energy business.

A second direction is making AI infrastructure itself a more flexible consumer of power, treating data centres as assets that can flex their consumption in response to grid conditions, in the same way industrial demand response has worked for decades in other sectors, but applied to compute workloads that can in principle be shifted in time or location. Vendors are now building this flexibility directly into reference designs for AI facilities, with claims of unlocking meaningful amounts of otherwise stranded grid capacity if facilities are designed to operate this way from the outset.

A third, longer-horizon direction is direct investment in new generation capacity, including a renewed wave of interest in nuclear and fusion power specifically framed around AI demand, with some of the largest funding rounds of the past year going to companies pursuing fusion power with AI infrastructure customers explicitly in mind.

For organisations planning AI infrastructure at any scale below the gigawatt campuses making headlines, the practical lesson is the same in miniature. Power availability, not just at the site but at the rack and row level, needs to be part of the initial design conversation, not a constraint discovered after the compute has been specified. Liquid cooling, power density planning and electrical infrastructure are no longer back-of-house considerations that can be handled separately from the compute and networking design. They are part of the same design problem, and increasingly the part that determines how quickly a facility can actually go live.