CHR in a PJM Heat Wave

On July 3, PJM published an operations update with a sentence that caught my attention:

“PJM issued a warning to prepare some transmission owners and utilities for the possibility of curtailing data centers and other large loads in their regions and moving them to backup generation. Ultimately, that action was not required and service was not impacted.”

So basically the largest power grid in the United States, serving 67 million people, spent Fourth of July week pushing past a demand record that had stood for 20 years. Wholesale prices in northern Virginia, home to the largest concentration of data centers on Earth, blew past $2,500 per megawatt hour. The Department of Energy had pre-authorized PJM to force data centers onto backup generation as a last resort before rolling blackouts.

The data centers ran straight through all of it. Nobody had to order them to do anything, and no price the market produced gave them a reason to leave.

Who actually flinched

Somebody did leave the system last week: from what I could find, it looked like roughly 6,000 megawatts during the July 2 evening peak. So this was contracted demand response: customers paid in advance to accept mandatory curtailment during emergencies. Factories, commercial buildings, and industrial users who signed up to be the load that leaves when the grid gets tight.

So the flexibility on the system was purchased ahead of time from the participants with the lowest tolerance for high prices. The load class at the center of every reliability debate, the one PJM’s own forecasts say will drive nearly all demand growth through 2030, contributed nothing. It was never asked, and at $2,500 per megawatt hour it had no economic reason to volunteer.

I made this argument in March in my piece The Flexibility Illusion: data center demand response will not materialize organically as economic behavior, because the price tolerance of AI compute sits so far above wholesale price ranges that curtailment destroys more value than it saves. The hyperscaler does not curtail; it pays someone else to.

July was that argument playing out in real time, documented in PJM’s own operations updates.

The CHR Price Impact

$2,500 per megawatt hour sounds intense, and to a grid operator or to most industrial consumers, it is.

To frontier AI inference, it is roughly 5 percent of their price tolerance ceiling.

The Compute Heat Rate (CHR)1 framework measures the maximum electricity price an AI workload can rationally pay before the computation becomes uneconomic. For frontier inference workloads, that ceiling currently sits above $50,000 per megawatt hour; blended across workload types, it runs in the thousands. This is not to be confused with a forecast of where prices will settle, instead this is a measurement of where the demand-side brake engages, based on the economics.

This means the worst pricing event on America’s largest grid in 20 years did not get within an order of magnitude of the point where frontier AI compute has an economic reason to blink.

Data centers didn’t cause the emergency, weather did, but this was layered on top of congestion and thin reserves.

The Compute Heat Rate is sometimes misread as a prediction that electricity prices are going to $50,000. Instead, it measures a ceiling: the price at which a given class of demand rationally exits. Below that ceiling, the load stays.

For a century, the implicit ceiling for large industrial load sat low enough that scarcity pricing worked. Prices spiked, demand left, the system self-corrected. That feedback loop is the unexamined assumption underneath wholesale market design, reliability planning, and most of the policy debate about data centers.

I’ve been saying for a while that when reserve margins get thin (which they’re projected to be across the country by 2030 at the latest), this type of thing is more likely to occur, more often. So this won’t be the last data point. The summer is yet young.

Hans Royal is the originator of the Compute Heat Rate™ (CHR) framework. All views are his own and do not represent those of any employer or affiliated organization.