Data Centre Energy: The Load That Arrived Faster Than the Wires

Why the Load Was Flat for a Decade

Between 2010 and 2020 global data centre workloads grew by roughly a factor of ten while their electricity consumption rose by only a few percent. That decoupling is one of the more striking efficiency achievements of the period, and it had three causes.

The first was the migration from small server rooms to hyperscale facilities. A cupboard with four servers in an office runs at low utilisation with poor cooling; a purpose-built facility runs its hardware at high utilisation with engineered airflow. Consolidation alone delivered large savings, and it happened for commercial reasons rather than environmental ones.

The second was virtualisation. A physical server running one application typically idles at 10 to 20 percent utilisation while drawing more than half its peak power, because an idle processor is not a free processor. Running many virtual machines on one physical host raises utilisation and eliminates most of that waste.

The third was Moore's law doing what it did: performance per watt improved steadily, so each new hardware generation delivered more computing for the same electricity. Together these absorbed a tenfold workload increase, which is why data centre energy stayed off the policy agenda for a decade.

All three are now largely exhausted. Consolidation has happened, virtualisation is universal, and performance-per-watt gains have slowed as transistor scaling approaches physical limits. The efficiency buffer that absorbed growth is gone, and growth now shows up directly as electricity demand.

PUE and What It Stopped Measuring

Power usage effectiveness is the standard metric: total facility electricity divided by the electricity that reaches the computing equipment. A PUE of 2.0 means every watt of computing costs another watt of cooling, lighting and power conversion. A PUE of 1.1 means the overhead is 10 percent.

Industry averages fell from around 2.5 in 2007 to roughly 1.5 today, and the best hyperscale facilities operate at 1.1 or below. That is a genuine achievement and it is also close to a floor: at 1.1 the overhead is already small, so even perfecting it would save 10 percent of a number that is itself growing rapidly.

This is why attention has shifted from PUE to the computing load itself. Free cooling - using outside air directly when the climate allows - drove much of the improvement, which is why facilities cluster in Ireland, the Nordics and the northern United States. Hot-aisle containment, higher operating temperatures and liquid cooling delivered the rest.

PUE also says nothing about whether the computing was useful. A facility running at PUE 1.1 while executing wasteful code is efficient by the metric and wasteful in reality, and no widely adopted metric captures useful work per joule - partly because nobody agrees what useful work is.

What AI Changed

A traditional server rack draws 5 to 10 kilowatts. A rack of AI accelerators draws 40 to 130. That is not an incremental change; it is a different building.

Air cannot remove that much heat from that small a volume, so liquid cooling becomes mandatory rather than optional - either cold plates on the chips or immersion of the whole board in dielectric fluid. Electrical distribution, floor loading and the physical layout all change with it. Facilities designed before 2022 frequently cannot host AI hardware at all without substantial rebuilding.

The workload profile changes too. Training a large model runs flat out for weeks, which looks to a grid like a steady industrial load. Inference - answering queries - varies with user demand and can be geographically shifted to where power is cheap or clean at that hour. That flexibility makes some AI load a candidate for demand response, and several operators have begun offering it.

Published projections of AI electricity demand vary by more than a factor of five, which is a reasonable signal of how uncertain they are. They depend on assumptions about model efficiency, hardware improvement and how much inference actually gets used - and efficiency per operation has historically improved faster than most forecasters assumed, which is the same pattern that made the 2010s projections too high.

The Constraint Is Local

A modern hyperscale campus can draw several hundred megawatts, comparable to a small city, at a single point. National generation can supply that. The substation cannot.

The mismatch is one of timescales. A data centre can be designed, permitted and built in 18 to 24 months. A new transmission connection at that scale takes 5 to 10 years, most of it in permitting rather than construction. The result is queues: Dublin and parts of northern Virginia and Amsterdam have at times refused new connections outright, not for lack of electricity but for lack of wires.

This has made data centre operators into unusual energy customers. They sign long-term power purchase agreements that finance new wind and solar farms, they are among the largest corporate buyers of clean energy in the world, and several have contracted directly for nuclear output including restarting retired reactors - because what they need is large, firm, round-the-clock supply at a single location, which is an unusual requirement that only a few sources satisfy.

The honest framing is that data centres are a large and fast-growing load whose absolute size is still modest beside industry, heating or transport, and whose local concentration creates real problems that national statistics hide. The debate would be clearer if those two facts were stated together more often than they are.