Unlike chip supply, grid capacity is bought in a public auction, and the clearing price is published. That auction has now cleared at its regulatory ceiling three years running and still failed to buy enough power to meet the region’s reliability standard. That makes it the most legible constraint in the AI buildout, and not the one that gets argued about. My claim is the narrow one: not that chips are easy, but that this shortage is already on a public scoreboard.

The market has already priced it

PJM coordinates wholesale electricity in “all or parts of” thirteen states and the District of Columbia. Its capacity auction is a forward market: generators bid to guarantee they will be available, and the clearing price is what the region pays for the assurance that supply meets peak demand. By design it runs “three years before it is needed”, which is meant to give developers time to build against the signal.

It has not been running on that schedule. The last three auctions cleared 10, 17 and 23 months before their delivery years began, not 36, so the signal reaches developers with a third to two thirds of the intended lead time. The gap is narrowing, but it has not closed.

Read the last four results in order.

• 2025/2026: $269.92/MW-day across the RTO, PJM’s whole footprint, with the BGE and Dominion zones clearing far higher at $466.35 and $444.26.
• 2026/2027: $329.17/MW-day, at the FERC-approved cap, for the entire footprint.
• 2027/2028: $333.44/MW-day, at the cap again.
• 2028/2029: $325/MW-day, at the cap a third time.

The last three are not prices the market discovered. They are the ceiling, and the ceiling itself moved: PJM calls the 2028/2029 result a “2.5% decrease from the 2027/2028 Base Residual Auction cap of $333.44/MW-day.” The fall from $333.44 to $325 is an administrative parameter changing, not demand easing.

Which is why the prices are not the strongest evidence here. This is: both of the last two auctions bought less than the region needs. The 2027/2028 auction came up 6,623 MW short of the one-event-in-ten-years reliability standard. The next one came up 6,831 MW short. PJM’s own framing is that these were “the first in PJM history in which the entire RTO fell short of the reliability requirement.”

That is what a capped market looks like when it cannot clear: price stops carrying the signal at the ceiling, so the shortage comes out as quantity instead. PJM’s executives are blunt about the cause. In December 2025, Stu Bresler, then PJM’s EVP for market services and strategy, said “data centers’ demand for electricity continues to far outstrip new supply.” In July 2026, CEO David Mills put it as “demand for electricity continues to grow faster than electricity supply.”

Two queues, stacked

Supply is responding. It just has to pass through two waiting lines that run in series.

The first is the interconnection queue. In April 2026, PJM closed the application window on the first cycle of its reformed process and reported 811 projects requesting 220 GW of nameplate capacity. Mills attributed the volume to demand “driven by data centers, advanced manufacturing and broader electrification trends.” PJM describes the review as “designed to be a one- to two-year process”. That is the study. Financing, permitting and construction come after it.

The second queue is for the machines themselves. Of that 220 GW, 105.8 GW is natural gas, the largest single category. GE Vernova reported in July that its “Gas Power equipment backlog and slot reservation agreements grew from 100 to 116 GW”, and that it remains “on track to deliver 20 GW of annual gas turbine output in the third quarter of 2026, with 24 GW in 2028.”

Divide one by the other. At 20 GW a year, a 116 GW book is 5.8 years of production. And the gas capacity requested in a single cycle in one grid region equals 91% of that entire global order book.

Read that 116 GW carefully: GE Vernova does not split it. It combines firm backlog with slot reservation agreements, and a reserved slot is an option on a build position rather than an order. Some will lapse. What it establishes is the queue for factory capacity, not a guarantee of delivery. The company also expects the same measure to reach “at least 125 GW by year-end 2026”, which lengthens the queue and shrinks that 91% at the same time.

I am not claiming those developers all intend to buy from one manufacturer. It is a scale check: one region’s appetite in one cycle is the size of one major manufacturer’s entire book.

What it costs

Take a one-gigawatt data center and price only its capacity obligation at the most recent clearing price.

1,000 MW  ×  $325/MW-day  ×  365 days  =  $118.6 million per year

Treat that as a floor, deliberately. It uses the lowest of the last three clearing prices and excludes energy, transmission and the reserve-margin multiplier that makes the actual invoice larger. Every simplification pushes the number down.

Run the same arithmetic on the 2024/2025 price, $28.92/MW-day, and the identical facility owed $10.6 million. That line item grew elevenfold in four auctions - roughly the window in which the current buildout was underwritten.

$118.6 million a year is not a rounding error against a facility’s economics. More to the point, it is a transfer. Capacity payments do not build anything by themselves; they move money from load to whoever already owns dispatchable generation in the right place. The AI operator books a cost. The incumbent generator books revenue.

There is a second cost that never appears on an invoice. The semiconductor industry settled this argument long ago by splitting equipment time into states, and the expensive ones are where a machine is fine but not producing. A data center that is built, financed and staffed but waiting for its grid connection to be switched on is exactly that: capital in a non-productive state, depreciating on schedule, earning nothing.

Where I could be wrong

A price at the cap tells you the cap binds, not how scarce power physically is. Reading a capped price as pure physics overstates my case. The quantity shortfall is the harder evidence, and it is the number I would watch.

Behind-the-meter generation routes around all of this. An operator that co-locates its own turbines or contracts directly with a nuclear plant is not buying capacity in this auction. If that becomes the dominant pattern, the auction stops describing the AI industry’s actual cost of power.

A queue is not a build. Most projects that request interconnection never reach commercial operation, and 220 GW requested is emphatically not 220 GW arriving. That cuts both ways. Less supply is coming than the queue implies, which sharpens the shortage. But it also makes 105.8 GW an upper bound on real turbine demand, and my comparison against GE Vernova’s book is generous to my own argument by exactly that much.

And the constraint is being actively worked. In June 2026, FERC approved PJM’s Expedited Interconnection Track, a “temporary, standalone process to quickly advance projects of significant size to address the urgent need for more capacity resources.” It takes up to ten requests a year at 250 MW or more, targets an interconnection agreement in ten months and operation within three years, and sunsets at the end of 2027. GE Vernova, for its part, is “implementing actions to produce 30 GW in 2030.” Both are real, and both take years to show up.

The number that would settle this is one nobody publishes: of an operator’s announced capacity, how much has a signed interconnection agreement and a date to switch on. Until then, the capacity auction is the closest thing to a public scoreboard, and it has been printing the same result for three years.