Behind-the-meter power: Why AI data centers are building their own gas plants

Behind-the-meter power means a data center produces some or all of its own electricity on its own site, instead of drawing everything from the public grid. The name comes from the utility meter: power made on the customer’s side of it never passes through the grid’s billing point. AI companies are turning to it because grid connections are slow.

Why a data center would skip the grid queue

Imagine opening a factory that needs as much electricity as a small city. Before you can switch on, the local utility must build power lines and substations, and that can take years. Many AI developers do not want to wait that long, because every month without compute is a month without a product.

Behind-the-meter (BTM) generation is the shortcut: build a power plant next to the buildings. Power Engineering, citing an Enverus forecast, describes data centers skipping the grid queue and turning to on-site power. The American Public Power Association, as summarized in the same coverage, identified 46 planned data centers with BTM generation, adding up to 56 gigawatts, roughly 30% of planned US data center capacity.

For scale, a gigawatt is a thousand megawatts, and a megawatt is roughly what a few hundred homes use at a given moment. A single large AI campus can need hundreds of megawatts or more.

What is actually on the site

Gas turbines are the most common choice. They burn natural gas to spin a generator, much like a jet engine bolted to a dynamo. Few other technologies deliver this much power in a compact footprint on a short timeline.

Common building blocks include:

• Aeroderivative turbines: compact turbines adapted from aircraft engine designs, built for fast installation.
• Reciprocating engines: large piston engines, similar in principle to a car engine at industrial scale.
• Fuel cells: devices that convert fuel to electricity chemically, pitched as lower-emission and easier to permit.
• Batteries: storage that smooths sudden changes in demand.

Many designs deliberately overbuild. Several smaller units mean one failure does not take the site down, according to the Power Engineering coverage.

Not every project is fully disconnected. Some sites run on their own generation permanently, some use it as a bridge until the utility finishes its upgrades, and some keep a grid link and use the on-site plant to cover the gap or to ease peaks. The label covers a spectrum, not one design.

A concrete example: Crusoe in Abilene

GE Vernova, a turbine maker, says Crusoe ordered 10 of its LM2500XPRESS aeroderivative turbines for a data center outside Abilene, Texas, then acquired 19 more, for 29 in total. GE Vernova says they are expected to provide nearly 1 gigawatt. Each unit is about 35 megawatts, which the company equates to 11 or 12 diesel generator sets.

GE Vernova, which sells the equipment, also claims fast start and stop, a smaller footprint and the ability to work almost like a battery. Its article describes the setup as grid-connected but not grid-dependent. Those are vendor claims, not independent measurements.

The example shows why the model appeals. Instead of one giant machine, the site uses many mid-sized units that can be installed in stages and switched on as the buildings come online.

Where it came from and who uses it now

On-site generation is old: hospitals, factories and mining sites have long run their own plants, and data centers have always kept diesel generators for emergencies. What is new is the scale and who is doing it. The generators are no longer only a backup; in some projects they are the main supply.

The International Energy Agency’s 2025 Energy and AI report puts data center use at about 415 terawatt-hours in 2024, roughly 1.5% of global electricity, and projects a base case of around 945 TWh by 2030. Coverage of the report notes many US developers are turning to on-site gas, often paired with batteries.

Examples are spreading. Power Engineering reports xAI used mobile turbines and reciprocating engines at its Memphis operation as a bridge while grid substations were built. And the Financial Times reported that Anthropic struck a $3.4bn deal for gas-fired power. We covered that report in Anthropic reportedly spends $3.4bn to build its own gas power for data centers. The counterparty, capacity and sites are not public, so it is not known whether that project is behind-the-meter in the strict sense.

The trade-offs and limits

Speed has costs. Turbines are scarce: Power Engineering reports that large-frame models from many major manufacturers are sold out into 2030 and prices have risen. That helps explain why a buyer might bypass usual suppliers, as the FT’s summary of the Anthropic deal describes.

Other limits:

• Emissions and permits: burning gas releases pollutants and carbon dioxide. xAI’s Memphis permit for up to 247 MW across fifteen turbines was challenged by the Southern Environmental Law Center, which alleged regulatory breaches. That is an allegation, and air-quality permitting is a growing constraint.
• Cost and fuel supply: the plant needs a steady gas pipeline, and the owner takes on operations that utilities normally handle, from maintenance to fuel contracts.
• Reliability: a site without a grid link cannot lean on it, so it needs spare units or batteries.
• Local acceptance: neighbors tend to care about noise, air quality and who pays for shared infrastructure, and a visible gas plant makes those questions concrete.

There is also a trade-off between speed and permanence. A bridge plant built to beat a grid delay may stay in service long after the grid arrives, which raises the question of what the plant is for once the wait is over.

What to watch next

Hybrids are one trend: on-site gas combined with batteries and flexible demand. The IEA’s summary also points to flexible operation of on-site generation and storage as a way to ease grid risks. Watch for fuel cells as lower-permitting-risk options and for whether projects later connect to the grid and sell power back.

Also watch disclosure. If AI labs fund power plants directly, investors and neighbors will want to know where they are, how big they are and what they emit. Market-research forecasts of rapid growth, such as the 2026 to 2032 projection cited in coverage, should be treated cautiously.

Finally, watch equipment supply. As long as turbines are sold out for years, the companies that secured them early hold an advantage, and deals that go around traditional suppliers may become a regular feature of how AI infrastructure gets built.