The Load Curve

Data centre electricity consumption is set to more than double by 2030 while staying under 3 percent of world demand. The aggregate is the least useful number in the dataset. This report works through the concentration, the workload mix and the supply chain constraint that the headline conceals.

The aggregate understates the constraint

Global data centre electricity consumption was around 415 terawatt hours in 2024, roughly 1.5 percent of world electricity consumption, having grown about 12 percent a year over the preceding five years. The IEA base case has it more than doubling to about 945 terawatt hours by 2030, slightly more than the entire current electricity consumption of Japan, and reaching just under 3 percent of the global total.

Three percent of anything rarely justifies a report. We are writing one because the aggregate is the least informative number in the dataset, and it is the number that gets quoted.

Regional concentration

The United States adds around 240 terawatt hours between 2024 and 2030, an increase of about 130 percent on its 2024 level. China adds around 175, up about 170 percent. Europe adds more than 45 and Japan around 15. Together the two largest markets account for close to 80 percent of global growth.

Within those countries the concentration goes further, because data centres site where land, fibre and interconnection allow rather than where demand is spread. A national figure of three percent is compatible with a regional figure many times that, and it is regional grids that have to be planned, financed and paid for.

A load that is three percent of a continent and thirty percent of a county is not a three percent problem. It is a county problem with national financing.

The IEA puts this in a form worth repeating: data centres are projected to account for more than 20 percent of total electricity demand growth across advanced economies through 2030. Growth is where system costs are set. In the United States, data centres are expected to consume more electricity for processing data by 2030 than all energy-intensive manufacturing combined, including aluminium, steel, cement and chemicals.

The composition of the load

The headline growth rate hides two very different trajectories. Electricity consumption in accelerated servers, the component driven by artificial intelligence workloads, is projected to grow around 30 percent a year. Conventional servers grow at 9 percent. Accelerated servers account for almost half the net increase in global data centre consumption to 2030; conventional servers for only around 20 percent, with the remainder in other IT equipment, cooling and infrastructure.

This matters for two reasons. First, it means the forecast is really a forecast about one workload type, and its error bars are the error bars on AI adoption rather than on internet usage. Second, accelerated computing has a different load profile: denser, hotter, and with different implications for cooling and for the ratio of infrastructure power to IT power.

Total global data centre capacity has reached almost 100 gigawatts. That figure is a better guide to the physical planning problem than any consumption number, because capacity is what has to be interconnected.

A constraint that is not electricity

Power is the constraint that gets attention. It is not the only one. Materials demand runs alongside it: gallium, used in computing chips and power electronics, is expected to reach more than 10 percent of current supply by 2030 on the IEA's projections. China accounts for around 99 percent of global refined gallium supply. That is a single point of failure with a policy dimension, and delays in a constrained input show up as delays in delivered capacity rather than as a visible price.

Per capita consumption gives a sense of how unevenly this is distributed. Africa consumed less than 1 kilowatt hour of data centre electricity per person in 2024, rising to slightly under 2 by the end of the decade. The infrastructure of the AI economy is being built in a small number of places.

What we are least sure of

• The efficiency path. Historical data centre energy forecasts have systematically overshot because efficiency improved faster than assumed. If that repeats, the base case is too high.
• Adoption. Because accelerated servers account for almost half the net increase, the projection is highly sensitive to a single behavioural assumption about AI uptake.
• Delivery. Interconnection queues, turbine and transformer lead times and construction labour may bind before capital does. In that case the load arrives later and more slowly than the demand forecast implies, which is a timing error rather than a level error.

Our reading is that the level is roughly right and the timing is optimistic. We would treat the 2030 figure as a reasonable central case and the path to it as considerably lumpier than a smooth doubling.

Source register