When companies used to site a data center, they looked first at connectivity, land and taxes. Power was assumed to be there. Today, the first question on more and more projects is: when can we get power?
Using public data, this piece tries to answer three questions: how fast demand is growing, where the bottleneck is, and what on-site generation changes.
Demand: growing fast, and concentrated
The International Energy Agency (IEA) estimates that data centers used about 415 terawatt-hours of electricity in 2024, roughly 1.5% of global consumption, growing by about 12% a year since 2017. In its base case, that figure more than doubles to around 945 TWh by 2030, with AI as the main driver.1
The US numbers are even more striking. A report by Lawrence Berkeley National Laboratory (LBNL) found that US data centers used 176 TWh in 2023, or 4.4% of the country’s electricity. By 2028 that could reach 325 to 580 TWh, or 6.7% to 12%.2
Harder than the total is the concentration. The IEA notes that an AI data center can draw as much electricity as an aluminum smelter, and that data centers are far more geographically concentrated: nearly half of US data center capacity sits in five regional clusters.1 For a regional grid, that means connecting a huge load in a small area in a short time.
The bottleneck: grids are built in years
Compute is built on a monthly rhythm: GPUs arrive, racks go up, systems are commissioned. Fast projects finish in a few months. Grids move on a yearly rhythm.
Take the US. New power plants first queue for interconnection studies. LBNL’s Queued Up report shows that the typical time from interconnection request to commercial operation rose from under two years for projects built in 2000–2007 to over four years for those built in 2018–2023, with a median of five years for projects built in 2023.3 Large new loads also wait for substations and transmission lines to be expanded, on a timeline measured in years.
This is not only a US problem. The IEA estimates that unless grid risks are addressed, around 20% of planned data center projects worldwide could be delayed. In advanced economies a new transmission line takes four to eight years to build, and wait times for critical grid equipment such as transformers and cables have doubled in the past three years.1
Compute delivered in months on one side, grid capacity added over years on the other. The gap in between is where on-site generation comes in.
The answer: generate first, wait for the grid later
A telling example comes from the utility American Electric Power (AEP). In November 2024, AEP announced an agreement securing up to 1 GW of Bloom Energy solid oxide fuel cells (SOFCs), the largest utility fuel cell initiative in the US. AEP was blunt about the purpose: to let data centers and other large customers get power quickly while the grid is built out. The fuel cells sit at the customer’s site, are designed not to send power back to the grid, and are paid for entirely by the customer under a special contract; they run on natural gas at first, and the technology could use hydrogen later.4
In June 2025, the Public Utilities Commission of Ohio approved the first project under that agreement: on-site fuel cells at two data centers operated by AWS and Cologix in central Ohio, with the two companies covering the full cost.5
On-site generation, in other words, is no longer just backup power. It has become a parallel supply path alongside the grid: bring compute online with on-site power first, then switch to or combine with the grid once it arrives.
Each on-site option has trade-offs
| Option | Strengths | Main constraints |
|---|---|---|
| Gas turbines / reciprocating engines | Mature, large units | Emissions, permits, fuel supply; gas turbine deliveries take several years1 |
| Fuel cells (e.g. SOFC) | Modular, compact, efficient | Mostly natural gas today, so not zero-carbon |
| Batteries | Fast response, smooth loads, shave peaks | Good for balancing, not for long-term baseload |
| Diesel generators | Reliable, the traditional backup | Emissions and permits limit long running hours |
| Nuclear (incl. small reactors) | Stable and low-carbon over the long run | Longest build time; a longer-term option |
The right choice depends on whether a project needs the fastest start, the lowest long-run cost or the lowest carbon. You rarely get all three.
What it means for infrastructure
The power architecture changes. It used to be “grid + diesel backup + UPS”. More and more it becomes a microgrid of on-site generation, the grid and storage, with more complex controls and protection.
Power quality gets harder. AI training loads swing widely, and on-site sources lack the inertia of a large grid to absorb those swings, so UPS systems and batteries take on more smoothing and protection work.
Decisions move earlier. The power plan has to be settled at site selection and project approval. Whoever can bring a credible energy solution at that stage gets into the project first.
Three things to keep in perspective
- An announcement is not an operating plant. Separate “announced”, “contracted”, “under construction” and “operating” when reading the news; many big headlines stop at step one.
- Gas-fired power is not zero-carbon. Fuel cells and gas turbines can deliver power fast, but their emissions need their own accounting.
- What on-site power really buys is time. Its cost per kilowatt-hour is not necessarily lower; its value is the revenue from bringing compute online years earlier.
How this looks from China
China is the world’s second-largest market for data center electricity. The IEA estimates that China accounted for about 25% of global data center consumption in 2024, behind the US at 45%, and that it is the second-largest source of growth to 2030 after the US.1
The path will differ. Grid build-out, the energy mix and project approvals all work differently, so the US approach of running on gas first and waiting for the grid cannot simply be copied. One change worth watching: in May 2025 the National Development and Reform Commission and the National Energy Administration issued a notice on green-power direct supply. Wind, solar, biomass and other renewable plants may now skip the public grid and supply a single user through a dedicated line, and new loads may be paired with their own renewable projects.6 The policy is not aimed at data centers specifically, but it gives large loads a formal route to organize power on the customer side.
Whatever the market, one thing is shared: where the power is, and when it can arrive, is becoming the first question any compute build-out has to answer.
Sources
All figures come from the public sources below, published between April 2024 and June 2025.
Footnotes
-
International Energy Agency, Energy and AI, Executive summary, April 2025. https://www.iea.org/reports/energy-and-ai/executive-summary ↩ ↩2 ↩3 ↩4 ↩5
-
U.S. Department of Energy, “DOE Releases New Report Evaluating Increase in Electricity Demand from Data Centers” (LBNL, 2024 United States Data Center Energy Usage Report), December 2024. https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers ↩
-
Lawrence Berkeley National Laboratory, Queued Up: 2024 Edition, April 2024. https://emp.lbl.gov/publications/queued-2024-edition-characteristics ↩
-
American Electric Power, “AEP Leveraging Fuel Cell Technology to Power Data Center Growth”, 14 November 2024. https://www.aep.com/news/stories/view/9866/AEP-Leveraging-Fuel-Cell-Technology-to-Power-Data-Center-Growth/ ↩
-
American Electric Power, “PUCO Approves Onsite Power Project for Data Centers”, 5 June 2025. https://www.aep.com/news/stories/view/10262/ ↩
-
National Development and Reform Commission and National Energy Administration, Notice on Promoting the Orderly Development of Green-Power Direct Supply (Fa Gai Neng Yuan [2025] No. 650), 21 May 2025 (in Chinese). https://www.ndrc.gov.cn/xxgk/zcfb/tz/202505/t20250530_1398138.html ↩