Demand-Side Data Center Dispatch Revolution

PJM just told us data centers are the demand response the grid needs. But only if we let prices go high enough to activate it.

On May 6, PJM published “Powering Reliability Through Market Design,” a 70-page, CEO-signed white paper laying out three paths for reforming the largest wholesale electricity market in North America. The paper has already generated significant commentary around capacity market design, mandatory hedging, and differential reliability frameworks.

You can wonk out on the full paper here:

On pages 49-50, under “The Hyperscale Paradigm Shift,” PJM specifically uses the Compute Heat Rate™ (CHR)¹ framework to propose something that has seemingly never appeared in an official market operator document before: using workload-level demand-side price elasticity as graduated demand response.

This is exciting, because it is potentially a fundamental reconceptualization of how the demand side can participate in wholesale electricity markets going forward. This could be a blueprint of how other markets could reform as well. Let’s dig in!

The Old Model: Demand Is Inelastic

For decades, wholesale electricity market design has operated on a simple assumption: demand doesn’t necessarily respond to price in real time, for the most part. Residential customers don’t see real-time prices; commercial and industrial load can be on fixed-rate contracts. The grid balances supply and demand mostly by dispatching generators: turning on more expensive plants as demand rises, turning them off as it falls.

This is economic dispatch, and it operates exclusively on the supply side.

The demand side of the market has basically been treated as a block of undifferentiated consumption. Traditional demand response programs exist, but they’re binary: a factory either curtails or it doesn’t, usually in exchange for a modest payment that reflects the avoided cost of a peaker plant. The economic signal is simple: are you willing to turn off for $500/MWh? Yes or no.

This model worked when all load had roughly the same economic character. A steel mill, a data center, and a paper plant all consumed electricity at similar price tolerances and similar operational constraints. There was no reason to differentiate.

That world is over.

The New Reality: Data Center Demand Is Heterogeneous

AI data center workloads do not behave like traditional industrial load. This is not a difference of degree; it is a difference of kind.

A commodity inference workload (running a chatbot, serving basic API calls) generates equivalent revenue per MWh of electricity consumed in the range of $800-1,270. A mid-tier training workload generates revenue per MWh in the range of $8,000. A frontier inference workload (serving a state-of-the-art model at scale, like GPT 5.5 or Claude Opus 4.7) generates revenue per MWh in excess of $50,000.

This metric, the Compute Heat Rate (CHR), is the demand-side analog to the gas heat rate that generators use to determine their marginal cost of production. It tells you the maximum electricity price at which a given AI workload remains profitable.

PJM recognized this in their paper and used the CHR framework to look at full workload tier decomposition. Their conclusion was striking: this graduated price tolerance is “precisely the graduated, price-elastic demand response the grid needs.”

PJM is not saying data centers are a problem only- they are saying data centers are a potential solution that needs to be activated.

What Activation Looks Like

Here is the mechanism PJM is describing:

When wholesale electricity prices rise during a scarcity event, different data center workloads hit their economic breaking points at different thresholds. At $1,000/MWh, few workloads curtail; most are still deeply profitable. At $2,000/MWh, the lowest-value commodity workloads begin approaching their economic shutdown threshold. At $10,000/MWh, commodity workloads may curtail or virtually “migrate” to cheaper regions, releasing capacity back to the grid. At $50,000/MWh, only frontier inference continues operating.

This is a really cool solution: it’s a graduated, price-elastic demand-side dispatch that actually could help the grid tremendously over time.

Each workload tier responds to price independently, based on its own “CHR”. The grid gets exactly the flexibility it needs: low-value workloads shed first, high-value workloads persist longest, and the total demand response is proportional to the severity of the scarcity event.

PJM’s paper explicitly connects this to their Reserve Certainty Senior Task Force (RCSTF) reform, stating that the current ORDC pricing at $1,000-1,900/MWh “begins to create the signal necessary to activate this flexibility; a higher cap ceiling would reach further into the CHR distribution.”

Translation: the demand-side dispatch mechanism already exists in the economics. The market design just needs to let prices rise high enough to activate it.

The Paradox

This leads to PJM’s insight called the Scarcity Revenue Paradox:

The data center workloads that continue operating at the highest scarcity prices are the ones generating the most scarcity revenue for generators. A frontier inference workload running at $10,000/MWh is paying generators ten times the normal clearing price. That revenue is precisely what solves the “missing money” problem that has plagued capacity markets for two decades.

As PJM puts it: “the influx of the very loads contributing to the resource adequacy problem today will help to shrink the relative scale of the missing money problem over time.”

The loads causing the problem could also be part of the solution. But only if scarcity prices are allowed to rise high enough for the mechanism to work.

Why This Hasn’t Happened Yet

Three structural barriers prevent the demand-side dispatch mechanism from activating:

First, scarcity price caps are too low. Most ISOs cap energy prices at $1,000-2,000/MWh. At those levels, no data center workload curtails, because even the lowest-CHR commodity workloads are profitable well above that range. The price signal never reaches the threshold where graduated demand response begins. PJM’s RCSTF reform is explicitly designed to address this.

Second, the political economy resists high prices. PJM’s paper names this directly: the “credibility trap.” High scarcity prices are economically necessary to activate demand-side dispatch and solve the missing money problem. But high prices trigger political intervention, price caps, and regulatory backlash that suppress the very signal the market needs. So this is indeed and issue that needs to be dealt with; but if more supply doesn’t get the signal to build, then the supply demand imbalance will last longer either way.

What This Means

PJM has laid out three reform paths (mandatory hedging, differential reliability, and energy market transition). All three require a framework for understanding which loads curtail at which prices. All three depend on the demand side becoming a legible, measurable input to market design, rather than an undifferentiated block of consumption.

This is the beginning of demand-side dispatch for electricity markets. Not the simple, binary demand response programs of the past, but a graduated, workload-level, price-elastic system that treats demand as a heterogeneous economic input, just as supply has been treated for decades.

The supply side of electricity markets has been disaggregated, measured, and optimized for 25 years. The demand side could get the same treatment. And the metric that makes it legible is the Compute Heat Rate.

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.

1. Royal, Hans, The Compute Heat Rate: Quantifying AI-Driven Electricity Price Tolerance
and Its Implications for Wholesale Market Repricing (February 28, 2026).
Available at SSRN: http://dx.doi.org/10.2139/ssrn.6322318