Article — "Grid Operators Move to Curtail Data Centers First as Power Becomes the Binding Limit": Start with the number that should end the argument: 73% of the Cherokee Nation citizens surveyed by their own task force reportedly use generative AI tools, 34% of them daily. And 64% of that same population opposes hyperscale data centers on their reservation.
The report names the contradiction outright—"a clear disconnect"—and then does nothing to resolve it, because the disconnect isn't a misunderstanding to be corrected with fact sheets. It's a decision. People will use the output while refusing the physical plant that produces it.
That refusal is the whole story now, and it's arriving from two directions at once.
Last week this column tracked the consent wall: 500 jurisdictions passing moratoriums in a single month, national opposition to nearby data centers running 63% to 70%, New York's first statewide pause. This week the second wall went up, and it's harder, because it isn't a county commission you can rezone around. It's the grid operator itself.
Per Data Center Frontier, PJM Interconnection—the grid for 67 million people across 13 states and Washington, D.C.—filed its Reliability Backstop Procurement with FERC on July 31 and, in a separate August 13 filing, proposed a framework containing one sentence that does more work than the entire curtailment scheme around it. Beginning with the 2029/30 capacity auction, incremental new large loads that don't bring their own power supply will be excluded from the demand PJM procures for at all.
Read that slowly. Not curtailed as a last resort. Erased from the forecast. The grid stops promising to serve them.
For the entire history of the American grid, being a paying customer of sufficient size meant the operator planned to meet your load. PJM has just proposed to end that for data centers.
For the entire history of the American grid, being a paying customer of sufficient size meant the operator planned to meet your load. PJM has just proposed to end that for data centers. A 50MW-plus facility that shows up without its own generation is no longer a customer the grid must serve—it's a load the grid may shed.
The mechanics arrive in stages. Starting June 1, 2027, new large loads without their own capacity get curtailed first—before PJM deploys the Pre-Emergency Load Management programs that pay everyone else to cut back. A Large Load Registry tracks every 50MW site, its location, its ramp schedule, and whether it brought power. That registry becomes the map states use to decide who gets shed first.
The numbers behind this aren't subtle. PJM projects 70 GW of new large load by 2038 against roughly 15 GW of generation retired since 2022. Two consecutive capacity auctions have now failed to meet the reliability requirement—the 2028/29 auction cleared at its $325/MW-day price cap and still came up 6,831 MW short, something that had not happened before these two consecutive auctions.
The independent market monitor blames data centers for a 75.5% jump in regional power costs. PJM's board chair calls the trajectory "not sustainable" and says the "central affordability question is how the costs of that investment should be allocated."
That sentence is the tell. This filing isn't really about reliability engineering. It's about who pays for resource adequacy—and PJM just decided it won't be existing households. It will be the new loads themselves, or nobody.
Aurora Energy Research's Julia Hoos put it plainly: the plan pushes "the responsibility for procuring new generation onto the large loads themselves." Bring Your Own Capacity. The buildout must now carry its own power plant.
Here's where the first chokepoint feeds the second. A power plant takes time. Equipment lead times run three to five years because the factories are backlogged; permits take longer; and, as one analyst noted, neighborhoods sometimes fight the plant just as they fight the data center.
But curtailment starts June 2027. There is no version of a new gas or nuclear plant that clears permitting, procurement, and interconnection in the window that matters. The three-to-five-year clock and the June 2027 deadline do not touch.
So the near-term answer is the only thing that installs fast: diesel generators and batteries. Federal rules let diesel run up to 50 hours a year for demand response, 100 for emergencies. That's the bridge operators will actually build across.
And that bridge lands them straight back into the consent wall. This week Vantage Data Centers came under fire for coordinating with Virginia regulators to cast doubt on a report tying diesel backup at a single 96MW facility to tens of millions in annual health damages for its neighbors.
Watch the loop close. The grid says bring your own power. The fast option is diesel. The diesel is exactly what the community moratoriums cite—air quality, noise—as the reason to ban the thing in the first place. Cherokee Nation's task force named energy and water consumption, air quality, and noise as its citizens' top concerns.
Both walls are now saying the same three words: not automatically, here. The grid won't automatically plan for you. The ground won't automatically zone for you. Firm power and community consent have become the two things you must separately win, and both PJM and the Cherokee Nation just proved they can withhold them.
Compute does not stop when it hits a wall. It migrates. That is the geographic consequence, and Blake Hurst named it in three words in the Wall Street Journal: "Developers, make me an offer."
Hurst, who ran the Missouri Farm Bureau for a decade, is publicly volunteering his farm 40 miles from a $6.3 billion project that Nodaway County just halted. His argument is coldly correct: a blocked project doesn't vanish, it relocates while the tax base stays gone. In Mason County, Kentucky, a developer facing holdouts simply redrew its site plan around willing sellers and kept moving.
So compute flows toward jurisdictions willing to trade consent for tax base—the declining county with a 13% population drop and a shrinking university, chasing 130 jobs and a billion dollars over 35 years. That's one tier of the new map.
The other tier is the self-powered island. If the grid won't procure for you, you leave the grid. Bloom Energy reportedly projects onsite generation as the primary power source rising from 13% of US data centers in 2024 to 38% by 2030—and fully grid-independent facilities from 1% to 27%.
That 27% is the number that redraws everything. More than a quarter of data centers building their own power stations, bypassing the resource adequacy question entirely, because the grid operator invited them to.
But building your own generation before June 2027 is, by the analysts' own math, nearly impossible—and it costs a fortune only the deepest pockets carry. Amazon, Google, and Microsoft can afford to self-generate; the smaller and newer companies without deep pockets cannot. Microsoft restarted a Three Mile Island reactor to feed Virginia. That is not a strategy available to a startup.
So the two-tier map hardens. Where the grid is tight—Northern Virginia, the corridors through Ohio, Pennsylvania, Maryland—only the self-generating giants build, because only they can bring their own capacity to a grid that no longer promises to serve them.
Everyone else migrates to the willing county, offering itself up like Hurst, competing on how much consent it will trade for how much tax base. PJM can't flip the switch itself—it told Reuters it lacks the authority and needs states to carry out the cuts. But it doesn't need the switch. It only needed the forecast.
By striking unbacked loads from the demand it plans for, PJM didn't just change how the grid responds to shortage. It changed where compute is allowed to physically land—and handed the deciding vote to whoever controls firm power and local consent. Those two gates, not chips or capital, now draw the map.