How much electricity do AI data centers use?

In detail

The global number is small, and it is growing fast

The International Energy Agency estimates that data centers used about 415 TWh of electricity in 2024, about 1.5% of global consumption, and that their use has grown about 12% a year over the last five years. In its Base Case, that figure reaches about 945 TWh by 2030, just under 3% of the world’s electricity. These numbers cover every data center workload: email, streaming, storage and AI.

The AI share shows up in the growth. The agency projects that electricity use by accelerated servers, which it says is mainly driven by AI adoption, grows about 30% a year, compared with 9% for conventional servers. Accelerated servers account for almost half of the projected net increase. The United States and China account for nearly 80% of the growth to 2030. (IEA figures checked October 24, 2026.)

The strain is local

A 3% global share sounds manageable, and on a world scale the agency calls it limited. A data center does not draw from the world grid, though. It draws from one utility’s lines. The Lawrence Berkeley National Laboratory report released by the US Department of Energy puts US data center use at 58 TWh in 2014 and 176 TWh in 2023, about 4.4% of US electricity. It estimates 325 to 580 TWh by 2028, or about 6.7% to 12%.

Supply takes longer than demand. In its executive summary, the IEA notes that new transmission lines can take four to eight years in advanced economies, and that wait times for transformers and cables have doubled in the past three years. Jeetu Patel describes how the industry is responding: instead of pulling power to where a data center is wanted, companies now build data centers where power is available. He expects countries to want their own sovereign data centers as well.

Kyndryl’s Kris Lovejoy works on the same problem from the operator side. She says PJM, the organization that forecasts load for its grid region, comes back every year to say it underestimated the previous year’s demand, always by a double-digit percentage. In her view, newer facilities are efficient, with liquid cooling and recycling; the issue is the number of devices that need power.

Efficiency inside the building is not the main lever

Andy Masley, who fact-checks AI environmental claims, argues that data centers are already heavily energy-optimized, because their operators are highly incentivized to cut energy use. He thinks the useful argument is about whether a given facility should be built and whether the uses of AI are worth the energy. Pressuring companies to shave consumption inside the building matters much less in his view.

Lovejoy points to one inefficiency she sees in practice: power is provisioned into a data center and then goes unused, so the operator files another connection request for more. Better rack planning recovers some of that capacity without a new grid connection.

Separate your prompt from the buildout

The personal question and the grid question have different answers. A single chatbot prompt is a tiny amount of energy, which the explainer on whether using AI is bad for the environment works through with Masley’s comparisons. The data center buildout is where demand, siting and the power mix are decided. Masley makes the same split: once you are talking about data centers, you are talking about the grid, both demand and supply.

The IEA also models the other direction. It estimates that existing AI-led efficiency measures in buildings, if scaled up, could save about 300 TWh of electricity, roughly the annual generation of Australia and New Zealand combined. That is a projection of what is possible, not a measured result.

Where it falls short

Forecasts disagree widely. The LBNL range for 2028 spans 325 to 580 TWh, and the IEA’s 2035 cases run from about 700 to 1,700 TWh depending on efficiency gains and how fast AI is adopted. These headline totals cover all data center workloads; they are not standalone estimates of AI use. When you see a single number for “AI’s electricity use,” check what it counts and which year it projects.

Databricks’ Denny Lee is more concerned about energy limits than inference costs for the next few years: inference will keep getting cheaper, but energy is harder to solve and takes more time. That is a forecast, not a measurement. For hardware choices that change the energy cost per answer, continue with why AI runs on GPUs.

Disclosure: Episode 45 is sponsored by Galileo, the Season 3 presenting sponsor; the thanks is at 0:00. Episode 22 was sponsored by Galileo, and Conor notes Databricks’ investment in Galileo at 13:44. Conor worked at Galileo at the time.