Energy, where it stands now

This is the first of a nine-part series under Where each part of the AI pipeline stands now. The map of the whole chain, from the power plant to the person using the answer, is The nine parts of the AI pipeline. Energy is part 1 because every later part sits on electricity.

I am writing for someone who uses an AI tool and has not had to think about where its power comes from. The question here is where that electricity is as of mid-2026. Data centers run AI and the rest of the internet in the same buildings, so a data-center total is larger than AI alone. Where a source splits electricity use at AI-focused data centers out, I say so.

Two uses of the same electricity

A data center uses electricity in the computers, and more electricity to convert power, hold a backup and move heat out.

The International Energy Agency (IEA) uses a ratio called power usage effectiveness, or PUE, for that split. PUE is the site’s total electricity divided by the electricity used by the information technology (IT) equipment. A PUE of 1.2 means that for every 1.0 unit the IT equipment uses, the building uses 0.2 more on cooling, power equipment and lighting.

In Key Questions on Energy and AI, published April 16, 2026, the IEA says the best large cloud sites now reach a PUE of about 1.1 to 1.2. Older sites, and sites a single company runs for itself, are often around 1.5 to 1.7. Moving from 1.6 to 1.2 cuts the electricity used outside the IT equipment by about two-thirds, and cuts the site’s total electricity by about a quarter. The report’s point, which I share, is that much of that gain is fixed at design. Cooling layout, including liquid cooling, and the local climate are costly to redo once the site is up. The path from a power plant to a chip has four steps.

1. A power plant makes the electricity. A buyer can sign a contract for it.
2. The grid carries it to the site, which means a connection the grid operator has agreed to.
3. Inside the building, equipment converts the power, holds a backup and removes heat.
4. The chips draw power while they do the multiplications a model is made of.

Training and running the model both happen at the last step. Training is the large job of making a model. Running it, called inference, happens each time a person asks for an answer, for as long as the model stays in service. The same April 2026 report says the balance of AI-related electricity use has shifted from training to inference.

How much, as of 2026

Nobody has one meter on every data center. The figures below are published estimates. I date each one and name who published it. A world total from the IEA and a United States total from Lawrence Berkeley National Laboratory (LBNL) stay separate. They cover different places, and their past figures are for different years.

As of 2024, the IEA estimated that data centers used about 415 terawatt-hours (TWh), about 1.5% of the world’s electricity. The United States accounted for 45% of that use, China 25% and Europe 15%. Since 2017, data center electricity use grew about 12% a year, more than four times the growth rate of electricity use overall. That estimate is in the agency’s Energy and AI report, published April 10, 2025.

On April 16, 2026 the IEA updated that picture in the report above and in its news release the same day. In the central case, electricity use at AI-focused data centers triples between 2025 and 2030. Bottlenecks, the agency says, make the more aggressive near-term cases less likely.

U.S. data centers take a larger share of U.S. electricity than the 1.5% world share. LBNL’s 2024 United States Data Center Energy Usage Report, published December 19, 2024, estimated that U.S. data centers used 176 TWh in 2023, 4.4% of U.S. electricity. Use had sat near 60 TWh a year from 2014 to 2016. The report’s scenarios for 2028 run from roughly 325 to 580 TWh, about 6.7% to 12% of U.S. electricity. The Department of Energy’s note publishes the 2023 and 2028 figures.

LBNL’s follow-up, the 2025 update, is dated June 2026. Its reference case is 649 TWh in 2030, which it puts at 11.8% of U.S. electricity. Other cases in that update fall between 9.5% and 15.3%. When the laboratory stacks the uncertainties, the bounds are 521 to 843 TWh in 2030. LBNL describes the method as bottom-up: planned shipments of equipment, how much electricity each device uses, and simulations of cooling. Those 2030 figures are model output.

The 649 TWh U.S. reference case and the 950 TWh world case come from different models with different scope and inputs, one for the United States and one for the world. Both are for 2030. I leave them side by side.

Each task can get cheaper while the total rises

The April 16, 2026 report says electricity per AI task has fallen by at least tenfold a year in recent years, from changes in software and in hardware. A simple text query, it says, now typically uses less electricity than a television running for the same length of time. Heavier work can use hundreds or thousands of times more per query. The report’s examples are video generation, reasoning steps and tasks that run through many steps on their own.

On those same pages, major model providers reported about three times the active users and about five times the revenue over the prior year. The IEA says there is no comprehensive count of how often, or how deeply, people use AI, and that sentence does not name the providers. I treat those two multiples as the providers’ own reports.

Drawing on Epoch AI’s record of leading chips, the IEA says performance per watt rose about 30% to 40% a year on average over the past decade, unevenly, with gains slowing as transistor scaling matures. Heavier use can outrun that gain. Electricity use at AI-focused data centers was up 50% in 2025, a year when a plain text query had come to use little electricity. The IEA’s news release puts capital expenditure at five large technology companies above $400 billion in 2025, with another 75% rise expected in 2026. That spending is why so many projects are looking for power at once. It counts dollars, a different unit from the terawatt-hours above.

More power in the same rack

A server rack in an advanced data center is about the size of a large refrigerator. The IEA says that by 2027 a rack that size could draw, at peak, as much power as about 65 households. The report bases that figure on Nvidia’s announced Rubin chip design. From 2020 to 2025, the power density of AI servers rose 11 times. The agency expects about a further fourfold rise by 2027. Density here means how much power sits in a given space.

Heat follows the watts. The report says high-density racks increasingly need liquid cooling, because liquid carries heat better than air does. Delivering those watts also takes more transformers and power electronics, which the IEA lists among the tight supply lines.

Getting a connection is the slow step

A data center can be built in about one to three years. Planning, permitting and building new grid infrastructure can take five to fifteen years. That comparison is from the IEA’s Electricity 2026 chapter on grids. The April 2026 report repeats that a wait for a grid connection can run five to ten years in many places, that lead times for transformers average two to three years, and that deliveries of gas turbines can take about five years, a figure the IEA attributes to Wood Mackenzie (2025).

Texas makes the queue visible because the grid operator publishes it. The Electric Reliability Council of Texas (ERCOT) had a large-load connection queue of about 63 gigawatts (GW) in December 2024, and about three-quarters of that queue was data centers. By January 2026 the queue was over 230 GW. The state’s all-time peak demand is 85 GW. The IEA drew those figures from ERCOT’s own data. The chart’s caption says not every project in the queue gets built.

The same report describes four ways developers are lining up power.

• Renewable contractsIn 2025 the tech sector accounted for about 40% of the corporate renewable power purchase agreements signed worldwide. Renewable agreements signed by data center companies, the IEA says, equal almost half the sector's current electricity use.
• Nuclear agreementsConditional deals for data centers to take power from small modular reactors (SMRs) grew from about 25 GW at the end of 2024 to about 45 GW. The report's summary puts the 45 GW at the end of 2025. Its chapter says a further 20 GW was announced in January 2026 to reach that total. The IEA expects the first commercial SMRs around 2030.
• Onsite gasWhere a grid connection is slow, U.S. developers are advancing projects with gas generation on the site. About one in five of the projects the IEA tracked by satellite has started construction or site preparation. The 2030 range is about 15 to 27 GW, mostly in the United States. Supplying a critical, swinging load reliably from onsite gas means building 30% to 70% more generation than the demand. With turbines scarce, the IEA does not treat onsite gas as a clearly faster route at large scale. Most data centers still prefer the grid.
• BatteriesTraining runs and ordinary use make the power draw swing. The IEA's estimate is that around 20 to 25 GW of longer-duration battery storage could sit in data centers worldwide by 2030, and that those batteries could become an asset to the grid if the incentives are right. The report counts these separately from the backup batteries that carry servers through short outages.

My reading of the 2026 figures

My reading: I am leaving the rough doubling by 2030 as the IEA’s projection. Chips on a purchase order do not run until a building can feed them. This part of the pipeline will show up later as delay and, in places, as cost. Electricity is one limit I can point at with a date and a publisher. The other eight parts have limits of their own. A last piece in this series will set them next to each other and look for where the whole pipeline is stuck.

The short account of all nine parts, and the links as they are written, is Where each part of the AI pipeline stands now.