Seven companies will sign a voluntary pledge at the White House promising to bring their own power. Meanwhile, forty-six data centers are already building fifty-six gigawatts of behind-the-meter generation. Three-quarters of it runs on natural gas. The pledge is the politics. The grid is the physics.
Tomorrow, representatives from Amazon, Google, Meta, Microsoft, OpenAI, xAI, and Oracle will gather at the White House to sign the Rate Payer Protection Pledge — a voluntary commitment to 'build, bring, or buy' their own power supply for new AI data centers. The ceremony is scheduled. The press releases are drafted. The political optics are arranged.
But the physical response started before the political one. Across the United States, forty-six data center facilities are already building their own power generation — a combined fifty-six gigawatts of behind-the-meter capacity. Ninety percent of these projects were announced in 2025 alone. The companies did not wait for a pledge. They looked at the grid, did the math, and started building a second one.
The Second Grid
Behind-the-meter means exactly what it sounds like: power generation that sits on the data center's side of the utility meter, never touching the public grid. The data center generates its own electricity, consumes it on-site, and the utility never sees the load. From the grid operator's perspective, these facilities are invisible — massive consumers of fuel that draw zero watts from the system they are physically adjacent to.
Cleanview identified the technology breakdown for twenty-three gigawatts of these projects where specifications are public. Seventy-five percent is natural gas. Not solar. Not wind. Not nuclear. Gas turbines, modular generators, combined-cycle plants. The projects targeting delivery in 2026 and beyond rely almost exclusively on natural gas because it is dispatchable, reliable, and fast to deploy. A gas turbine can be installed in weeks. A solar farm takes months. A nuclear plant takes a decade.
This creates a specific kind of irony. The companies signing the Rate Payer Protection Pledge are, in aggregate, the same companies that have published net-zero commitments. Google pledged to run on carbon-free energy twenty-four hours a day, seven days a week, by 2030. Microsoft committed to being carbon-negative by 2030. Meta set a goal of net-zero emissions across its value chain by 2030. Amazon pledged net-zero by 2040.
The behind-the-meter buildout tells a different story. When faced with a choice between waiting for clean energy infrastructure and building gas turbines now, the companies chose gas. Not as a bridge technology — there is no public timeline for transitioning these plants to renewables. As the primary power source. The net-zero pledges are aspirational. The turbine orders are contractual.
What They Are Actually Building
The company-level details are more revealing than the aggregate.
In El Paso, Texas, Meta is constructing a one-gigawatt data center campus. To power it, Meta contracted Enchanted Rock to install eight hundred and thirteen modular natural gas generators — a four-hundred-and-thirty-seven to four-hundred-and-seventy-three million dollar facility that will operate exclusively for Meta for five years before any grid connection is considered. Eight hundred and thirteen generators is not a backup system. It is a private power plant disguised as distributed generation.
Google took a different path. In December 2025, Alphabet acquired Intersect Power for four and three-quarter billion dollars. Intersect operates 2.2 gigawatts of solar capacity and 2.4 gigawatt-hours of battery storage, with several additional gigawatts in development. Google's plan is to co-locate data centers with generation — energy parks where the power plant and the server farm share a fence line. This is the cleanest version of the behind-the-meter model, but it required Google to become a power company, not partner with one.
Microsoft revived a nuclear power plant. The company signed an eight-hundred-and-thirty-five-megawatt power purchase agreement with Constellation Energy to restart Three Mile Island — the same facility whose partial meltdown in 1979 halted nuclear construction in the United States for a generation. The reactor had been shut down for economic reasons in 2019. Microsoft's demand made it economically viable again. The contract is for twenty years. A commitment that long is not a hedge. It is an infrastructure thesis.
Crusoe Energy is building at a different scale entirely. Project Jade in Cheyenne, Wyoming, will deliver 2.7 gigawatts using gas turbines and hydrogen fuel cells. In Abilene, Texas, the Stargate facility targets 1.2 gigawatts with a mid-2026 delivery date. GE Vernova is supplying twenty-nine LM2500XPRESS turbine packages — roughly one gigawatt of combined capacity — with units installable in as little as two weeks. The speed matters. Crusoe is not building for the decade. It is building for the quarter.
Bloom Energy projects that thirty-eight percent of data center facilities will use on-site generation by 2030, with twenty-seven percent running entirely on self-generated power. If those estimates hold, roughly a quarter of American data centers will operate as electrically independent installations — connected to the grid for backup but generating their own baseload. A parallel power system, purpose-built for computation.
What the Grid Operator Sees
The behind-the-meter buildout does not make the grid's problems disappear. It shifts them.
PJM Interconnection — the grid operator serving sixty-five million people across thirteen eastern states — held its capacity auction for the 2027-2028 delivery year. The clearing price hit three hundred and thirty-three dollars and forty-four cents per megawatt-day. The previous auction cleared at twenty-eight dollars and ninety-two cents. An eleven-fold increase in a single cycle.
Data centers accounted for six and a half billion dollars — forty percent — of the sixteen-point-four billion dollar total auction cost. Nearly 5,100 megawatts of demand increase traced directly to data center interconnection requests. Ratepayers across PJM's thirteen-state territory will collectively pay an additional one-point-four billion dollars in capacity costs starting June 2026. These are not generation costs, which the pledge addresses. They are capacity costs — the price of maintaining enough reserve power to keep the grid reliable when demand spikes. The pledge says nothing about capacity markets.
In Texas, the ERCOT interconnection queue exploded from sixty-three gigawatts in December 2024 to two hundred and twenty-six gigawatts by November 2025 — a fourfold increase in eleven months. Data centers represent seventy-three to seventy-seven percent of large-load interconnection requests, or roughly one hundred and sixty-five to one hundred and seventy-four gigawatts. Over half of queue requests lack completed interconnection studies. The queue is not a forecast of what will be built. It is a measure of what has been requested — and the volume is so far beyond the system's capacity to process that the queue itself has become a bottleneck.
The grid operators see what the pledge obscures. Even when data centers generate their own power, they still need grid interconnection for redundancy, ramp-up periods, and maintenance windows. They still drive transmission investment. They still affect capacity markets. The behind-the-meter model reduces the megawatt-hours a data center draws from the grid. It does not reduce the megawatts the grid must be prepared to deliver.
Three Hundred Bills
While the federal government prepares a voluntary pledge, the states are doing something different. They are legislating.
In the first six weeks of 2026, over three hundred bills were introduced across more than thirty states targeting data center power costs, siting, and community impact. Five new moratoriums were enacted in the past month alone — Michigan, Indiana, North Carolina, and two Colorado jurisdictions. Thirty-nine data center projects have been formally rejected by local authorities since 2023.
Virginia — which hosts the largest concentration of data centers in the world — introduced SB 253, proposing a rate rebalancing: data center electricity rates would increase 15.8 percent while residential rates would decrease 3.4 percent. The state's utility commission has already approved requiring large-load customers to pay eighty-five percent of distribution and transmission demand charges and sixty percent of generation demand charges, effective January 2027. Virginia is not banning data centers. It is repricing them.
The state response reveals something the federal pledge cannot. Electricity rates are set at the state level, by utility commissions, through rate cases — not by White House ceremonies. A Harvard Law energy scholar put it plainly: the White House is putting this pledge on the wrong entities. The companies signing tomorrow do not set electricity rates. They never did. The utilities and commissions that do set rates are not signing anything.
The Physics of the Bet
This journal has tracked the AI infrastructure buildout across nine entries. The Foundation documented the six-hundred-and-fifty-billion-dollar capital commitment. The Hard Hat found that the real bottleneck was electricians, not balance sheets. The Lot Line discovered sixty-four billion dollars in blocked projects. The Meter measured the demand curve. The Pledge analyzed the political theater of a voluntary commitment that addresses thirty percent of the cost chain.
The grid adds the physical dimension. Fifty-six gigawatts of behind-the-meter generation is not a policy response. It is an engineering one. The companies looked at the public grid — its age, its capacity constraints, its regulatory complexity, its political vulnerability — and decided to build around it. Not to improve it. Not to partner with it. To bypass it.
The implications extend beyond data centers. When the most capitalized companies in history build private power infrastructure at utility scale, they create a two-tier energy system. The public grid serves households and small businesses. The private grid serves computation. The public grid ages and underinvests. The private grid attracts capital at rates the public system cannot match. The divergence reinforces itself.
Data centers consumed about 4.4 percent of total U.S. electricity in 2023. Lawrence Berkeley National Laboratory projects that share will reach six-point-seven to twelve percent by 2028. AI-specific power demand alone — separate from traditional data center workloads — is projected to reach one hundred and sixty-five to three hundred and twenty-six terawatt-hours per year by 2028. That range's upper bound exceeds all electricity currently used by all U.S. data centers for all purposes today.
The pledge will be signed tomorrow. The grid is already being built. The politics acknowledge the problem. The physics reveal what the companies actually believe about the solution — which is that the public grid cannot deliver what they need, at the speed they need it, and they would rather spend billions building a parallel system than wait for the one that exists to catch up.
Eight hundred and thirteen gas generators in El Paso. A restarted nuclear plant in Pennsylvania. A four-and-three-quarter-billion-dollar solar acquisition. Twenty-nine turbine packages installable in two weeks. This is not a pledge. It is a grid.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)