How much electricity does AI actually use?
The honest answer has two halves that point in opposite directions. Globally, AI is a small share of electricity consumption and will stay a small share through 2030. Locally, it is already the largest new load on some grids in decades, and that is where the fights are.
The numbers that hold up
The most rigorous public baseline is the International Energy Agency’s Energy and AI report, published in April 2025.
It put global data center electricity consumption at around 415 terawatt hours in 2024, roughly 1.5 percent of world electricity use. That figure covers all data centers, including everything unrelated to AI: streaming, storage, email, ordinary cloud computing.
The base case projects that roughly doubling to around 945 TWh by 2030, just under 3 percent of global consumption. Growth runs at about 15 percent a year, more than four times faster than electricity demand from every other sector combined. Inside that total the split matters: accelerated servers, meaning the AI hardware, are projected to grow around 30 percent a year, while conventional servers grow about 9 percent.
Since that report, the trend has stayed in the upper part of its range rather than the lower. Data center electricity demand grew about 17 percent in 2025, with consumption at AI-focused facilities rising roughly 50 percent, and 2025 total consumption came in near 485 TWh.
For anyone reading a scary number in a headline, the useful check is which of three things it measures: all data centers, AI data centers specifically, or a single model’s training run. Those differ by orders of magnitude and get quoted interchangeably.
Why 3 percent still breaks grids
A global percentage is the wrong instrument for a load that concentrates.
Data centers are built where land is cheap, fiber is dense and interconnection is available, which means a handful of counties absorb an enormous share. Nationally, US data centers sit around 4 to 5 percent of consumption today, and projections for 2030 run from 9 to 17 percent depending on assumptions. In the specific counties hosting hyperscale campuses, the local share is far higher than either figure.
The binding constraint is not generation in aggregate. It is interconnection: the queue to connect new load to the transmission network, the substations and lines to carry it, and the local generation to firm it. Those take years to build. A data center takes months. That mismatch, not a shortage of electricity in the world, is what produces moratoriums and county-level bans, and it is why Musk’s warning that AI would run out of electricity before it ran out of chips moved from a provocation to something close to industry consensus during 2026.
The politics that follow from this are covered separately in why AI data centers face local backlash. The measurement point here is narrower: any figure quoted as a share of global electricity describes a market that does not exist, because nobody buys power globally.
The numbers to distrust
Three categories of estimate circulate freely and are worth less than they appear.
Per-query energy figures. The energy cost of one prompt depends on which model answered, how long the answer was, whether the model reasoned before responding, which hardware generation ran it and how efficient the facility is. Labs disclose almost none of that. Any single number for “one AI query” is an assumption stack presented as a measurement.
Training-run totals used as ongoing cost. Training a frontier model is a large one-time expenditure. Inference, meaning everyday use, is the recurring one, and at current usage it dominates the lifetime energy cost of a deployed model. A headline about the electricity used to train a model says little about what running it costs.
Straight-line extrapolations. Efficiency per unit of computation has improved substantially and continues to. Model efficiency has improved too. Both cut against the extrapolation, and neither cancels growth in demand. The IEA’s own range across scenarios is wide for exactly this reason, which is a reason to quote the range rather than the base case alone.
What is not in dispute is direction. Demand is growing faster than any other electricity sector, it is concentrated where the grid is least able to absorb it, and the response so far has been local restriction rather than national planning.
Related coverage
- Musk’s warning that AI will run out of electricity is becoming consensus, how power replaced chips as the constraint.
- Meta disclosed $279 billion in future data center leases, what the build-out costs.
- DeepSeek is building a gigawatt of compute in Inner Mongolia, the same demand outside the US grid.
- SpaceX is recruiting engineers to build AI clusters in orbit, how far the constraint pushes the proposals.
- Why are AI data centers facing local backlash?, what communities do once the load arrives.
- AI datacenter money tracker, every committed dollar as it lands.
Quick answers
How much electricity do data centers use?
The International Energy Agency estimated global data center consumption at around 415 terawatt hours in 2024, roughly 1.5 percent of world electricity consumption. That covers all data centers, not only those running AI workloads.
How much will AI electricity use grow by 2030?
The IEA's Energy and AI report, published in April 2025, projects data center consumption roughly doubling to around 945 TWh by 2030 in its base case, just under 3 percent of global electricity. That is about 15 percent annual growth, more than four times the growth rate of all other electricity demand combined. Accelerated servers, the AI-specific hardware, are projected to grow around 30 percent a year against 9 percent for conventional servers.
Why do local communities fight data centers if AI is only 3 percent of electricity?
Because the load is not spread evenly. Data centers cluster where land, fiber and power interconnection are cheap, so a single county can absorb a large share of regional capacity. The constraint that bites is the interconnection queue and local generation, not the global percentage, which is why restrictions are written county by county rather than nationally.
How much electricity does one ChatGPT query use?
Per-query figures circulate widely and deserve skepticism. They depend on the model, the length of the answer, whether reasoning is used, hardware generation and data center efficiency, and the labs publish very little of that. Aggregate reporting from utilities and grid operators is the more reliable basis, which is why the credible estimates are built from facility-level power draw rather than from arithmetic on a single prompt.