When the Power Won’t Show Up, Do You Build Your Own?

Data Centers & Energy By EnergyInnoCore Editor · Published Oct 19, 2026 · Updated Oct 21, 2026 · 10 min read

For most of the data center industry's history, electricity was treated almost like water: you found a good site, negotiated with the local utility, built a substation, installed backup generators, and eventually the power showed up. That assumption is getting harder to make. The AI boom is changing the scale of demand so quickly that in some of the busiest markets, land and permits can be easier to secure than the megawatts needed to turn the servers on. A data center can be designed in months and built in a couple of years; new transmission lines and large power plants take much longer.

That mismatch raises an uncomfortable question for the industry: if the grid cannot deliver power fast enough, should AI data centers simply build their own? A few years ago that would have sounded extreme. Today it is part of normal project-development conversations, and grid operators themselves are starting to encourage it. This article looks at why, which technologies are realistic, and what it means for everyone else on the grid.

The Problem Has Shifted from Energy to Capacity

The United States is not running out of electricity in any general sense. The harder problem is getting enough of it to the right place, at the right time, with enough transmission and generation behind it. AI makes that especially difficult because an AI data center is not an ordinary commercial building. Proposed campuses often require hundreds of megawatts, and the largest clusters under discussion reach into gigawatt territory; at full load, a 1 GW campus uses electricity on the scale of a major city.

These facilities also arrive in clusters. In Northern Virginia, Texas, Ohio, Georgia, Arizona and other markets, utilities are not dealing with one large customer but with several developers asking for hundreds or thousands of megawatts in the same area. EPRI has estimated that U.S. data centers could use roughly 4.6% to 9.1% of national electricity generation by 2030, and that demand is heavily concentrated geographically. PJM, which runs the grid and power market for all or parts of 13 states and Washington, D.C., expects about 30 GW of load growth from data centers between 2024 and 2030. Thirty gigawatts is not a rounding error; it is a major infrastructure challenge.

Waiting Five to Seven Years Doesn't Work in AI

The AI industry runs on a completely different clock from the utility industry. Chip generations turn over quickly, and AI companies are competing to train larger models, deploy more inference capacity and win customers now, not in the mid-2030s. If a utility tells a developer, "We can give you 100 MW initially, but the remaining 600 MW needs transmission upgrades that won't be ready for several years," the economics of that site can change overnight. A billion-dollar server investment sitting in an unfinished building generates little value.

That is why power availability now influences site selection as much as tax incentives, fiber or land prices. Bloom Energy's 2026 data center power research describes power availability as having shifted from a planning consideration to a fundamental constraint on growth, and its survey found that on-site generation is increasingly part of long-term power strategy rather than a temporary fix. The distinction matters. Historically, data centers installed generators because the grid might fail. Now some developers are installing generation because the grid may not arrive in time in the first place.

"Bring Your Own Generation" Is Becoming Policy

The clearest sign of the shift came from PJM. In January 2026, as it set out how to integrate very large new loads while protecting reliability for its roughly 67 million customers, PJM endorsed a pathway for large customers to bring their own new generation, often called BYONG (later framed as "bring your own new capacity"), along with an expedited interconnection track for projects that pair new load with new supply.

Ten years ago, that would have sounded very unusual. The conversation used to be "Here is your utility connection." Increasingly it is "How much power can you bring with you?" That does not mean data centers become electric utilities. In many cases the new generation will still connect to the regional grid, operate through a utility, or be developed by an independent power producer. Economically, though, the data center customer is increasingly responsible for causing new generation to be built.

What Would a Self-Powered AI Data Center Use?

There is no single answer, and the industry is moving toward combinations of technologies rather than betting on one source. The U.S. Department of Energy has outlined a range of resources that can help meet data center demand; the main candidates for on-site or dedicated supply are below.

Natural gas: the practical near-term option

Whatever one thinks of it environmentally, natural gas is probably the most practical near-term option for many large campuses. Gas turbines and reciprocating engines are commercially proven, dispatchable and able to run continuously. Unlike solar and wind, they do not depend on the weather, and unlike large nuclear projects they do not require a decade-long development process. A developer may be able to build hundreds of megawatts of on-site gas generation faster than a utility can finish some major transmission upgrades.

The problem is carbon. The companies building AI infrastructure are also among the most vocal corporate supporters of emissions reduction, and large fleets of gas generators beside AI campuses make those commitments harder to meet. Recent analyses of planned U.S. data centers show utilities meeting part of the new demand with new gas plants and delayed fossil retirements. Gas can solve the speed problem; it does not necessarily solve the emissions problem. (Our guide to behind-the-meter strategy covers the permitting and fuel-supply issues in detail.)

Fuel cells: more relevant than many expected

Fuel cells are modular, relatively compact and able to produce electricity continuously on site. A developer does not have to build one 500 MW station on day one; modules can be added as new data halls come online, which fits how large AI campuses are typically built. Capital is following the idea: in June 2026, Brookfield expanded its financing partnership with Bloom Energy to as much as $25 billion for power projects serving AI infrastructure and other large users.

Fuel cells are not a magic solution. Running on natural gas, they still emit carbon, and fuel supply, cost, manufacturing capacity and long-term maintenance all matter. But when the choice is between waiting years for a transmission upgrade and installing modular power next to the servers, the economics look different. We explore this trade-off in Fuel Cells and AI Data Centers.

Nuclear: the long game

From a purely operational perspective, nuclear power is close to ideal for AI: data centers want very large amounts of electricity 24 hours a day, and nuclear plants produce exactly that. The problem is timing. A developer trying to energize a campus in 2028 cannot depend on a new reactor scheduled for the 2030s.

Even so, the largest technology companies are planning well beyond the next few years. Google has backed Kairos Power's advanced reactor program, whose first 50 MW project is planned to supply TVA's grid and support Google's data center demand; Kairos broke ground on its Hermes 2 project in Tennessee in April 2026. Meta went further, announcing in January 2026 agreements with Vistra, TerraPower and Oklo that could support up to 6.6 GW of new and existing nuclear capacity by 2035. These are not backup-generator contracts. They show technology companies beginning to influence what kinds of power plants America builds. (For more on existing reactors, see Nuclear Power's Second Act.)

Batteries: essential, but not power plants

Battery storage will be part of the answer, though batteries are sometimes credited with something they cannot do: a battery stores electricity, it does not create it. A large battery can reduce peak grid demand, ride through outages, balance renewable output or support a facility while the grid is constrained, all of which is extremely valuable.

PJM has highlighted the possibility of data centers combining batteries with workload management to cut grid demand quickly during periods of stress, shifting some computing in time while storage supports the facility. Imagine a campus that normally draws 500 MW. During a grid emergency it pauses lower-priority computing, discharges its batteries and temporarily reduces its grid draw to 350 MW. That is a very different picture from the traditional view of a data center as an inflexible block of demand. But eventually the batteries must recharge, and something still has to generate the electricity.

The Likely Future: A Mini Power System

For these reasons, I don't think the future is simply "grid power versus self-generation." It is probably both. The next generation of AI campuses may look like miniature power systems, with:

• a utility grid connection;
• dedicated natural gas generation;
• fuel cells;
• large battery storage;
• nearby solar or wind;
• long-term nuclear contracts; and
• software continuously deciding which source serves which load.

The grid would remain extremely important, but it would no longer be the only source. This approach also allows campuses to develop in phases. The utility might initially provide 150 MW; the developer adds 200 MW of on-site generation and 100 MW of batteries; several years later a transmission upgrade brings another 300 MW. Some of the early generation could then be retired, converted to lower-carbon fuels or kept for reliability. That is far more flexible than simply waiting for the grid.

The Fairness Question: Who Pays?

There is another side to this story. If a large new data center requires billions of dollars of transmission, substations and generation, residential customers understandably do not want those costs quietly appearing on their bills. The issue is politically sensitive in every major data center state, and it strengthens the case for some form of bring-your-own-power. If AI companies want very large amounts of electricity very quickly, it is reasonable to expect them to pay for the new supply they require. Meta, for example, has said its energy agreements are intended to add generation and that consumers should not bear the costs of its data center demand. The details will vary by state and utility, but the principle is gaining ground, as our analysis of the Ratepayer Protection Act shows. Economic growth from AI is valuable, and so is affordable electricity for everyone else; the two have to coexist.

The Data Center Business Is Becoming an Energy Business

This may be the biggest change of all. For years, technology companies bought electricity. Increasingly, they will have to develop it. That means understanding gas pipelines, transmission planning, power markets, capacity auctions, nuclear licensing, battery degradation, turbine procurement, environmental permitting and project finance, subjects that once belonged mostly to utilities and independent power producers.

That may be healthy. The companies creating the new demand have enormous balance sheets, strong credit and a powerful incentive to build new supply. Rather than viewing AI only as a burden on the grid, AI demand could become one of the largest catalysts for new power infrastructure in decades, but only if the industry abandons the idea that electricity automatically appears at the property line. The next great AI data center may still be defined by its GPUs, fiber and servers. Increasingly, though, the first question developers ask will be simpler: where will the power come from? And if the grid's answer is "not for another five years," the industry's response is more and more likely to be: then we will build some ourselves.

EnergyInnoCore Editor

Independent analysis of power, grid and policy for AI data centers. Every article lists its sources. About the editor · Editorial policy

This article is for general information only and is not engineering, legal, financial or investment advice. See the disclaimer.