Behind-the-meter power for data centers, explained: how bring-your-own-generation works

Behind-the-meter power means a data center draws electricity from generation on its own site, ahead of the utility's meter, instead of buying everything from the grid. It is the quiet center of a loud argument. On October 21, 2026, President Trump said new large data centers are "mandated" to make their own electricity, and the engineering and rules behind that phrase are far more detailed than a social media post.

What "behind the meter" means

A utility meter marks where the public grid ends and a customer's private system begins. Generation on the customer's side of that meter is behind the meter. For a data center, that can mean gas turbines, fuel cells, batteries or a nearby power plant that feeds the campus directly. The facility can still keep a grid connection for backup and balancing, which is what makes the arrangement more than an island.

Developers are pursuing this increasingly. Bisnow reports that Meta's Hyperion campus in Louisiana is being co-developed with power plants. The motive is speed and certainty. A new load can wait years for grid studies, and own generation shortens the line.

Why the grid cannot absorb it fast enough

The bottleneck is the interconnection queue. Nvidia's own blog on its DSX platform, published October 21, 2026, says more than $300 billion in equipment is backlogged and over 200 gigawatts of projects are waiting in U.S. interconnection queues. It states: "Power is now the biggest bottleneck in infrastructure." That is a vendor view, from a company that sells the chips drawing the power, but the queue figures describe a real wait.

We covered the same pinch from the grid side in our report on the AI data center interconnection bottleneck. Own generation is the workaround: if you bring your own power, you ask the grid for less.

How the rules work: PJM's "bring your own generation"

The clearest rulebook is the one for PJM, the grid operator for much of the mid-Atlantic and Midwest. A regulatory explainer by Relae on FERC's April 16, 2026 order (Docket ER26-1088-000) summarizes it. The order approved four interconnection paths and rejected two PJM proposals, and told PJM to refile by May 18, 2026. The four paths are:

• Sub-full-capacity interconnection service, using capacity a host generator does not need.
• Request acceleration for co-located loads that can show no significant network impact.
• Provisional Interconnection Service, an interim service during the full study.
• Surplus Interconnection Service, using unused capacity at an already-studied generator's connection point.

The explainer's own illustration shows how it can work: a 1,000 megawatt data center with a 900 megawatt on-site generator requests 100 megawatts from PJM. Those are the author's examples, not FERC figures. It also notes a typical arrangement covers about 90% of load on-site and 10% from the grid, and that PJM's queue holds about 800 proposed projects, roughly 220 gigawatts.

Where Washington stands

Trump's post on Truth Social said: "New and large, in some cases gigantic, Data Centers are mandated by me to make their own electricity." The Hill and Bisnow both report it is unclear whether executive action will follow. Bisnow adds that earlier in October the Federal Energy Regulatory Commission rejected efforts to require data centers to bring their own generation to connect to the country's largest grid system.

So the policy picture has two layers. The industry practice is already spreading. The legal mandate, as of October 22, 2026, is a statement without published text. Do not assume the first implies the second.

The catches

Self-supply is not free power. The Relae explainer lists unresolved issues: penalty rate design for unreserved use, rules for behind-the-meter applications that sit in a separate proceeding, and reliability questions for co-located loads paired with storage. It also gives an example in which non-firm power drops out in about 15% of summer peak hours, a reminder that a facility that leans on the grid at its worst moments still imposes a cost.

There is also the question of fuel and emissions. Own generation at this scale often means gas, and none of our sources quantify the climate effect of the new on-site plants, so we do not either. What the sources do say is that communities object on more than price. The Hill reports that a medium-sized data center can use up to about 110 million gallons of water a year for cooling, according to an Environmental and Energy Study Institute analysis.

Can software stretch the power?

Nvidia's DSX blog describes tools to run the grid and the AI factory as one system. It reports that Phaidra, using DSX tools, got about 10% more compute within the same power envelope, and that Emerald AI's modeling suggests it could unlock up to 100 gigawatts on the existing U.S. grid. Both are vendor-reported and modeled results, not audited ones, and we have not seen independent confirmation. They point to a second route besides new generation: use the existing wires more flexibly.

The hype check

The claim: bring-your-own-power solves the data center grid problem. The evidence shows it speeds connection and reduces grid draw, but the rules are still being written, the mandate is unpublished, and the best numbers come from vendors or illustrative examples. It is a real tool, not a cure.

Our take

Watch the language. "Behind the meter" sounds like independence, yet most designs keep a grid tie and a reliability role for the utility. For the superintelligence race, the practical point is that compute is now gated by electricity. The labs that secure power fastest, through their own plants, flexible loads or both, will get compute online sooner. What to watch is whether PJM's refiled rules, any FERC follow-up and a White House text actually line up.