PJM's capacity auction has cleared at its negotiated price cap three times in a row, driven mostly by data center load. PJM's own no-cap simulation says the cap moved $13.3 billion and 77 megawatts. What that means for the power bill behind AI inference.
On July 14, 2026, PJM published the results of its auction for electricity capacity in the year from June 2028 to May 2029, and the report said it in one line: "all prices cleared at the cap ($325.00)." A few pages later it described how the cap worked: "All offers below the floor were cleared, and all offers above the cap were not cleared."
It was the third auction in a row to end that way. PJM runs the grid for more than 67 million people across 13 states and the District of Columbia, and once a year it buys capacity for a future delivery year. The design calls for buying three years ahead, though a compressed schedule has held recent auctions closer to delivery; the 2026/27 auction ran in July 2025. Capacity is a promise that a power plant, or a customer willing to cut its use, will be available when demand peaks. For delivery in 2024/25 that promise cleared at $28.92 per megawatt-day. Then it cleared at $269.92, then at the cap of $329.17, then at the cap of $333.44, and now at the cap of $325.00.
Those three capped numbers are the same number. Each report states the cap as $256.75 per megawatt-day of installed capacity. The headline figures move only because PJM converts that ceiling into "unforced capacity" using how reliably its reference plant is expected to perform, and that accreditation changed between auctions. When the December 2025 release called its $333.44 "a slight increase (+1.3%)," it was describing an accounting change, not a market that had moved. The market has hit the same ceiling three times.
For anyone building AI, that ceiling matters in two ways. It is part of what inference costs in a region whose load growth PJM attributes mostly to data centers. And the reports that document it are an unusually clear lesson in what a price cap does and does not do.
A capacity auction normally ends where supply meets a demand curve PJM draws, which it calls the Variable Resource Requirement curve. The cap is not part of that design. On December 30, 2024, Pennsylvania's governor, Josh Shapiro, filed a complaint arguing that the maximum price for PJM's capacity auctions was "unjust and unreasonable." PJM settled. The agreement, filed with federal regulators in February 2025, set a floor of $138.25 and a cap of $256.75 per installed megawatt-day for two auctions. PJM's Independent Market Monitor noted that this was the first time since the capacity market began that its demand curve had "both a defined maximum price and a defined minimum price."
Then the settlement was renewed. On February 12, 2026, PJM's board asked to extend the collar through the auctions for 2028/29 and 2029/30. It cited feedback from "the White House National Energy Dominance Council together with the governors of our 13 states supporting the extension." Federal regulators accepted the extension on April 28, 2026. A price that began as a litigation settlement now has the backing of the White House and every governor in the footprint.
PJM's own release for the December 2025 auction names the reason. The forecast peak load for 2027/28 was about 5,250 megawatts higher than the year before, and "nearly 5,100 MW of that increase is attributable to data center demand." Stu Bresler, PJM's executive vice president, said the auction "leaves no doubt that data centers' demand for electricity continues to far outstrip new supply."
The Independent Market Monitor put a price on it for the 2026/27 auction. Total revenues were $16,124,370,889. Including 11,993 megawatts of existing and forecast data-center load added $7,271,197,971 to that, an 82.1 percent increase. With the uncapped demand curve, the monitor estimated, the data-center effect would have been $14,189,483,234, or 278.0 percent. The monitor's conclusion was plain: "data center load growth is the primary reason for recent and expected capacity market conditions."
One of its scenarios is striking. Without any data-center load, total revenues would have been $8,853,172,918, which the monitor says is "due to using the restricted VRR curve with a minimum price." Take the data centers out and the auction clears at the floor. Put them in and it clears at the cap. The entire range the settlement allows is the data-center effect. The monitor is careful to say its scenarios "are not forecasts or predictions," but as a picture of what is moving this market, it could not be clearer.
PJM's reliability standard asks for a 20 percent reserve margin above forecast peak. The 2026/27 auction met it by 139 megawatts. The 2027/28 auction was, in PJM's words, "the first auction in which the entire RTO, including Fixed Resource Requirement areas, fell short of the reliability requirement": about 6,600 megawatts short, with a reserve margin of 14.8 percent. The 2028/29 auction came in 6,831.3 megawatts short, and PJM's report notes that this "is the second BRA in a row where PJM has cleared more than one percentage point below" its reserve target.
That matters because of a rule in PJM's tariff. The current provision "requires a Reliability Backstop Auction to be conducted if three consecutive BRAs clear more than one percentage point below" the target. The next auction, for 2029/30, would be the third. PJM says it is working on an update to the backstop rules, so the trigger may change before it is pulled. As written, though, it is one auction away.
Here is the part of the record that surprised me. Every one of these reports includes a simulation of the same auction with no cap and no floor, run on the offers PJM actually received.
For 2026/27, the uncapped price would have been $388.57, and "the cleared capacity is identical to the cap and floor case." For 2028/29, the uncapped price would have been $554.72, and $776.69 in the Commonwealth Edison zone around Chicago. That would have cleared 138,394.6 megawatts instead of 138,317.8, which is 76.8 more, an increase of 0.056 percent. The total bill would have been $29.7 billion instead of $16.4 billion, or $13.3 billion more, and the reserve margin would still have been 14.4 percent, well short of the target.
So in the auction's own horizon, the cap moves money, not megawatts. The plants that exist clear either way. A price 70 percent higher would have bought 77 megawatts. The cap decides whether $13.3 billion goes from electricity customers to generators, and it does not change how much capacity there is.
The record adds an odder twist. Because the 2027/28 auction cleared 6,623.2 megawatts less than the reserve target, PJM estimates the total was "$0.8 billion less than it would have been if the RPM cleared enough capacity to meet the IRM." In a capped auction that comes up short, the shortage is where the savings show up.
PJM is careful not to overclaim here, and anyone citing it should be too. The simulation holds offers fixed, and PJM writes that it "does not know how offers may have changed if the temporary cap and floor was not implemented." Its own summary warns that "a price cap below the established VRR Curve can reduce the amount of investment, and therefore supply, in the PJM region." The board has said much the same about the whole design: a "one-year capacity commitment procured three years forward may not match the current investment climate that may be required to incentivize new generation." The cap may cost nothing in megawatts this year and a great deal in megawatts later. The reports can show the first and can only warn about the second.
Economists have been describing this situation since before most of the grid existed. In September 1946, Milton Friedman and George Stigler wrote a pamphlet about rent control called "Roofs or Ceilings?" Comparing the housing crunch after the San Francisco earthquake with the one after the war, they put it in one line: "In 1906 the rationing was done by higher rents. In 1946, the use of higher rents to ration housing has been made illegal by the imposition of rent ceilings, and the rationing is by chance and favoritism."
PJM's version has a twist the pamphlet did not need. In the short run the supply of capacity barely responds to price at all. Supply offered in 2027/28 rose by 956 megawatts, and new generation and uprates cleared just 774 megawatts. In 2028/29, new generation and uprates cleared 524.7 megawatts. So the ceiling is not creating a shortage the way a rent ceiling does. The shortage exists at any price the auction could pay.
That puts the rationing somewhere else, and PJM's documents say where. "Approximately 57 GW of projects have completed PJM's study process and have either signed or been offered generation interconnection agreements and are free to proceed to construction," the December release says. "Many of these projects, however, continue to be slowed or stopped by factors unrelated to PJM, including local opposition, state/local permitting delays, supply chain challenges or financing." PJM says it has processed more than 170,000 megawatts of new generation requests since 2023. Its Reliability Resource Initiative drew 11,000 megawatts of proposed projects, and the board has committed "to initiating a backstop procurement to procure new supply." New capacity is being allocated by queue position, permits, construction schedules and negotiated procurement, while the price sits at a ceiling set by settlement.
Developers have seen the other version of this market. In late 2023, AWS launched EC2 Capacity Blocks, which let you reserve GPU instances "for a future date," up to eight weeks ahead, and its pricing page says "reservation prices are updated regularly based on trends in supply and demand." That is a forward capacity market for compute with no cap: when GPUs are scarce, the price says so. PJM runs a forward capacity market for the electricity those GPUs draw, and its price stopped being allowed to say so three auctions ago.
Converting these prices into something a builder can hold takes arithmetic, and the arithmetic here is mine, not PJM's. A megawatt-day of capacity spread over 24 hours of a load that runs flat, as a data center roughly does, works out as follows:
| auction | capacity price ($/MW-day) | per MWh of flat load |
|---|---|---|
| 2024/25 | $28.92 | about $1.21 |
| 2027/28 (at the cap) | $333.44 | about $13.89 |
| 2028/29 (at the cap) | $325.00 | about $13.54 |
| 2028/29 with no cap (PJM simulation) | $554.72 | about $23.11 |
Treat this as an illustration, not anyone's invoice. PJM's own release says "not all load pays this clearing price because of the impact of self-supply and bilateral contract arrangements," and large customers pay through contracts and hedges that smooth these numbers out. It is also not a cost per token or per query. That would need a measured energy figure for your model on your hardware, which these reports do not contain. Still, the direction is not in doubt: the capacity component of a flat load's power cost in this region rose roughly elevenfold across four auctions, and the uncapped price PJM simulates is about 1.7 times the cap.
Three things follow for anyone planning compute in PJM territory, or buying it from someone who does.
Budget for the uncapped number. The collar is a negotiated arrangement for a set run of auctions. Federal regulators have accepted it through the 2029/30 auction, and the 2028/29 auction already ran under it. PJM's own simulation says what the market would clear at without it. If you are signing a multi-year hosting, colocation or power contract, find out how capacity costs pass through and what the contract assumes after the collar ends. The simulation, not the cap, is the honest planning figure.
Price your flexibility. In PJM, a customer's share of capacity costs, its peak load contribution, is set using the system's five coincident peak hours of the summer, and Constellation says these costs may account for 20 to 30 percent of a total electric bill. A training run, a batch inference queue or an eval sweep that can pause through those few hours lowers the very number its capacity obligation is computed from, which efficiency work alone does not target. Flexibility that the grid can count on is worth money here.
Ask about position, not price. When price is not rationing supply, the scarce thing is a place in line: an interconnection agreement, committed capacity, a site whose power is already built. The useful question for a provider is not what the power costs. It is whether the power exists yet, and who else is waiting for the same megawatts.
The July report's own summary says the cap did two things at once: every offer below the floor cleared, and every offer above the cap did not. Neither built a megawatt. Whatever does get built in PJM over the next few years will be allocated by the queue, the permit office and the negotiating table, and anyone planning to run inference there should plan on those terms.
Pausing a training run or a batch queue through the few peak hours is worth money only if you can show it happened. If agents schedule your jobs, Chain of Consciousness keeps a tamper-evident record of what each agent ran and when, so a claim that the queue stood down through the peaks can be checked afterward.
pip install chain-of-consciousness npm install chain-of-consciousness
Or start without installing anything: Hosted Chain of Consciousness.