The cap is the floor now.

Yesterday afternoon, while everyone was still arguing about New York’s moratorium, the country’s largest grid operator quietly published a more important number.

PJM runs the capacity auction for thirteen states and sixty-seven million people. It’s the market that pays power plants to exist, to be there on the hottest day in 2028 when everything is running at once. Yesterday that auction cleared at $325 per megawatt-day. That is the maximum price the rules allow, and it’s the third year in a row the auction has slammed into it. The bill is $16.4 billion, and it lands on consumer electric bills through the spring of 2029.

A price cap is supposed to be the emergency brake. When the market rests against it three years running, it stops being a ceiling and starts being the price. The cap is the floor now.

Here is the number inside the number, and it’s worse. This auction attracted 525 megawatts of genuinely new generation. Over the same period, the demand forecast grew by about 2,000 megawatts, driven, in PJM’s own words, by large data center loads. The auction came up 6,831 megawatts short of the grid’s own reliability target. The entire point of a capacity market is to summon new supply with high prices. The prices have been maxed out for three years and the supply is not coming. Sixty-seven million people are now paying monopoly prices for scarcity that the market structure was designed to prevent.

Yesterday morning I wrote that New York banned a building, not compute. By yesterday afternoon, PJM had made the same statement in a different language. The regulator said pause. The market said pay. Both of them are pointing at the same object: a form factor that demands hundreds of megawatts of new interconnection in places where new megawatts do not exist at any legal price.

The sophisticated money already understands this. On Monday, Blackstone, Apollo, and KKR put $5.34 billion into Williams’ behind-the-meter power projects, generation wired directly into data centers, skipping the public grid entirely. Think about what that transaction actually is. Three of the largest investors on earth paid five billion dollars to not stand in line. The interconnection queue has become so valuable to avoid that avoiding it is now an asset class.

And the demand side is not slowing down to let anyone catch up. IBM lost a quarter of its value yesterday, its worst day since the 1987 crash, because enterprise customers abruptly moved their budgets into AI servers, storage, and memory. The buildout is no longer just competing for land and water and electrons. It is eating the rest of the technology budget. That money is coming for capacity, and capacity has to live somewhere.

So follow the logic to where it actually leads. New gigawatt campuses need new gigawatts, and the auction just told you, for the third consecutive year, that new gigawatts are not showing up. But the grid is not uniformly empty. It’s lumpy. There are small blocks of existing, already-interconnected power all over the country: at substations, at industrial sites, at facilities that closed and left their service behind. Nobody bids those megawatts in a capacity auction. Nobody writes press releases about them. They just sit there, stranded, in two and five and ten megawatt pieces, waiting for a load that fits.

That’s the entire design brief for our machine. A complete 2 megawatt AI data center that arrives on three trucks, runs in 15 working days on any pad with power and fiber, consumes zero water in every operating mode, stays under the noise ordinance, and leaves without a trace when the economics say move. I described it in more detail yesterday. The machine doesn’t ask the grid to build anything. It’s sized to the power that already exists, in the increments it already exists in.

A gigawatt campus has to reshape the grid around itself: new plants, new lines, new rate cases, a decade of fights. A 2 megawatt machine reshapes nothing. It doesn’t appear in a capacity auction or a rate case or a moratorium hearing. It shows up on a pad, does its work, and pays for power like any other industrial customer. One is a negotiation with sixty-seven million ratepayers. The other is a delivery.

Three years at the cap is not a market signal anymore. It’s a verdict. The grid cannot price the monuments into existence.

Monuments bid in auctions. Machines just plug in.

Henri Francois is the co-founder of Borealis by Stilllife, a data center fabrication company building modular, zero-water AI infrastructure in Corona, California. henri@borealis.cool