Data centers stopped waiting for the grid. They're building their own gas plants instead.

The most important decision in artificial intelligence this month was not about a model. It was about a gas turbine. Earlier in August, a joint venture between the utility PPL and Blackstone's infrastructure arm said it had locked up around five gigawatts of gas turbines to power data centers in Pennsylvania. Five gigawatts is not a data center. It is a mid-sized country's worth of generation, bought not to feed the grid but to sit next to a building full of chips and feed it directly. That is the story, and it is a boring one, which is exactly why it matters.

For two years the AI industry told itself a clean-energy story: the data centers would run on solar, on wind, on small nuclear reactors arriving any day now. Follow the megawatts actually being contracted in 2026 and a different, less flattering picture appears. The industry is not waiting for clean power. It is not really waiting for the grid at all. It is buying gas, building it behind the meter, and in Washington it is asking for permission to leave the regulated system entirely.

The queue is the whole problem

Start with the constraint, because everything else follows from it. To connect a large new load to the American grid, you join an interconnection queue, and that queue is now measured in years — five or more in the busiest regions. A model trained on last year's chips cannot wait five years for a plug. So developers have found the exit: generate your own power, on your own land, and never join the queue. In the industry's language it is "behind the meter." In plain terms it means a private power plant, usually gas, that answers to the data center and not to the grid operator.

The numbers have moved fast. By some counts, developers have announced on the order of 100 gigawatts of behind-the-meter gas generation in the United States, with dozens of individual data-center projects planning to self-generate tens of gigawatts between them. A behind-the-meter plant can be delivering power in roughly 18 to 36 months. The grid connection it replaces can take twice that. When speed is the binding constraint, and for this industry it is, the slower cheaper option loses to the faster dearer one every time.

They are not buying cheap electricity. They are buying time — and paying a premium per kilowatt-hour to skip a queue everyone else still has to stand in.

Price the power, not the press release

Here is the part the announcements skip, and it is the only part I care about: what does this electricity actually cost per kilowatt-hour? Because behind-the-meter gas is not cheap power. A dedicated on-site gas plant, often running in simpler configurations to get built quickly, produces electricity that is generally more expensive per kilowatt-hour than the grid average it is bypassing — more expensive than a big combined-cycle plant sharing its output across a whole region, and far more expensive than the queued-up wind and solar the developer decided it could not wait for.

So the trade is not power for less money. It is power for more money, sooner. The data-center operator is paying a premium per kilowatt-hour to compress five years into two, and it can afford to, because the thing the power feeds — a rack of chips depreciating fast enough that a year of delay is its own catastrophe — makes almost any electricity price look small. When your product is that valuable per watt, you stop optimising for the cost of the watt and start optimising for the speed of it. That is a rational choice for the operator. It is a distortion for everyone sharing the same gas market and the same grid.

The bill that formalises the exit

The behind-the-meter boom is a workaround. A bill introduced in the Senate this year would turn it into a category. The Decentralized Access to Technology Alternatives Act — the DATA Act — would amend the Federal Power Act to create a new kind of entity: a "consumer-regulated electric utility," a privately financed generation system built to serve a new load, physically islanded from the grid, and exempt from the federal rules that govern everyone else — rate regulation, reliability standards, interconnection, transmission planning.

The logic is tidy: if you never touch the public grid, why should the public's utility regulator touch you? And there is a genuine argument that keeping these enormous loads off the shared system protects existing ratepayers from having to fund their transmission. But read the condition carefully. The exemption holds only so long as the system stays fully islanded; connect to the grid for backup or to sell spare power, and the status collapses and the rules return. That is the tell. The design assumes these plants are walled gardens, when the thing every operator actually wants is a wall with a door in it — islanded when it suits them, connected when it doesn't.

What the bill really does is remove the last reason to wait. Today, going off-grid means giving up the regulated system's protections and its backup. Codify the carve-out and the off-grid gas plant stops being a stopgap and becomes the plan. The regulated grid, the one that spent a century learning to spread costs and keep the lights on for everyone, becomes the option you choose only if you are too small to escape it.

Who is left holding the grid

This is where the cost reappears, because in energy it always reappears; it never vanishes, it just moves. When the largest, best-financed new loads on the continent build their own power and never join the grid, two bills land on everyone else.

• The gas bill. Turbine manufacturing is a global, constrained supply — a handful of firms, order books reportedly full for years. Every gigawatt of turbines a data center buys is a gigawatt someone else cannot, and the scramble pushes up prices and lead times for all new gas generation, including the plants that ordinary grids need.
• The grid bill. The transmission network is paid for by the people connected to it. If the biggest new customers self-supply and opt out, the fixed cost of that network is spread over a smaller base — which means the households and small businesses still attached to the grid quietly pay a larger share of it.

And there is a wider meter running, the one I always come back to. The turbines being bought up for data centers in Pennsylvania and Texas come off the same global production lines that a utility in Lagos or Dhaka or São Paulo depends on to add generation. When the richest industry on earth corners a limited supply of power equipment to run chips, it lengthens the queue for power meant to run cities. The cost of AI's electricity does not stay in the data center's budget. It leaks outward, into the price of a turbine and the wait for one, in places that were already short of both.

The number that decides it

Strip away the announcements and the argument comes down to one figure, the one I keep for everything: cost per kilowatt-hour, and who pays it. The data-center operator has decided that a higher cost per kilowatt-hour, paid to its own gas plant and delivered in two years, beats a lower one paid to a grid it would wait five years to join. On its own spreadsheet, that math is correct.

The trouble is that the spreadsheet is incomplete. The premium the operator pays per kilowatt-hour is visible and it is theirs. The costs they push outward — the dearer turbines, the thinner grid cost base, the longer wait for power somewhere with less of it — are real and land on people who never chose the trade. Behind-the-meter gas is not the clean-energy future the industry advertised, and it is not, when you count all the meters, the cheap one either. It is simply the fast one, and speed, as always, is the most expensive thing to buy. The only question the DATA Act really asks is whether we are going to write that bill into law, or keep pretending the power was free.