The Flexibility Illusion

Why Data Center Demand Response Doesn’t Change the Structural Math

Data center demand response is having a moment. Some hyperscalers are signing agreements with utilities to shift or reduce portions of their electricity consumption during peak hours. Industry groups are developing frameworks to value this flexibility as a grid resource. The narrative is appealing: data centers are not just massive consumers; they are flexible partners that can ease grid stress and protect ratepayers.

The flexibility is helpful, but the conclusion that it resolves the structural impact of data center demand on electricity markets is not. Hyperscalers signing these agreements is a symptom of the problem at hand, not a blanket solution.

Understanding “why” requires a framework, or a way to measure the depth of demand-side price inelasticity across different types of data center operators and compare it to the layer of flexibility being offered.

That framework is the Compute Heat Rate™ (CHR)¹.

What Is the Compute Heat Rate?

The Compute Heat Rate measures the maximum electricity price a data center workload can sustain before the computation becomes uneconomic. It is the demand-side analogue of the gas heat rate, which converts fuel cost into the marginal cost of electricity generation for gas-fired power plants.

Where the gas heat rate asks, “At what electricity price does generation become uneconomical?” the CHR asks, “At what electricity price does AI compute become uneconomical?”

The formula is straightforward and it yields a dollar-per-MWh ceiling: below that price, the operator continues consuming electricity; above it, rational economics imply curtailment or relocation.

The numbers are striking. Frontier AI inference workloads generate revenue equivalent to over $50,000 per MWh of electricity consumed. Mid-tier inference generates roughly $8,000/MWh. Even the blended average across all workload types is approximately $6,350/MWh, or about 127 times the gas heat rate benchmark of ~$50/MWh. Traditional industrial consumers, by contrast, curtail at $60 to $160/MWh.

Access the foundational research at computeheatrate.com.

Why Measuring CHR by Player Type Matters

The data center industry is not monolithic. Different operators have fundamentally different relationships with electricity price, flexibility, and risk. The CHR framework reveals why by stratifying the market into three distinct player types, each with a different price tolerance profile.

Player 1: The Vertically Integrated Hyperscaler

The largest cloud and AI companies own everything: the servers, the software, the workloads, the G/or TPUs, and increasingly the power supply chain. When one of these operators signs a demand response agreement, it is making a decision about which of its own workloads to defer during peak hours.

This is possible because the hyperscaler has full visibility into the economic value of every computation running across its fleet. At the top of the stack sit frontier inference workloads with CHR ceilings exceeding $50,000/MWh. These will never participate in demand response. In the middle sit production cloud services and enterprise AI applications with CHR ceilings in the thousands: operationally inflexible during the hours that matter. At the bottom sit batch training, model evaluation, and internal analytics: deferrable, lower-value work that can shift to off-peak hours without meaningful business impact.

The demand response offering draws from the bottom of this stack. But measuring it against the full load profile reveals its scale: a thin layer of low-CHR compute sitting on top of an enormous block of high-CHR, inelastic demand.

The hyperscaler is not curtailing because it has to. It is curtailing because it can afford to. That distinction changes how we should interpret demand response.

When an aluminum smelter participates in demand response at $80/MWh, it does so because the economics compel it. Every MWh consumed above that price destroys value. The smelter’s flexibility and its price sensitivity are the same thing.

When a hyperscaler curtails deferrable workloads, it is voluntarily shifting low-priority work while its core operations continue generating revenue at rates that make the prevailing electricity price irrelevant. A blended CHR of $6,350/MWh means the operator could absorb $500/MWh wholesale prices across its entire portfolio and remain wildly profitable. The demand response is not a signal of price sensitivity, it’s a byproduct of extraordinary profitability.

Player 2: The Colocation Operator

Lots of U.S. data center capacity is operated not by hyperscalers but by colocation providers under service level agreements guaranteeing 99.99% or higher uptime.

The colocation operator does not own the workloads and cannot distinguish between a frontier inference job generating $50,000/MWh in revenue and a legacy enterprise application generating $200/MWh. It cannot selectively curtail low-value workloads because it does not know which workloads are low-value.

When wholesale prices spike, the colocation operator faces a binary choice: keep every tenant running at full power, or breach the SLA. A colo operator will pay whatever the market clearing price is before it will breach an uptime guarantee and face liquidated damages, customer churn, and reputational destruction. This makes every megawatt of SLA-locked colocation capacity behave as fully price-inelastic demand, by contractual construction rather than workload economics.

This segment cannot replicate the hyperscaler’s demand response model. Measuring the CHR of the colocation segment reveals that not much of this capacity is available for demand response under current market structures.

Player 3: The Industrial Consumer Sharing the Grid

The aluminum smelter, the chemical plant, the steel mill, the paper manufacturer. These facilities share the grid with Players 1 and 2. They draw power from the same transmission system and pay capacity charges determined by the same auction results. But their CHR equivalents, the maximum electricity price they can sustain, range from $60 to $160/MWh.

Before the data center buildout, these industrial consumers existed in a market where demand response worked as designed. During scarcity events, prices rose, industrial load curtailed in order of price sensitivity, and the market rebalanced. The clearing price during scarcity was effectively capped by the level at which enough demand would voluntarily exit to restore balance.

The data center buildout may break this mechanism. Hundreds of megawatts of load with CHR ceilings 50 to 1,000 times above traditional industrial thresholds do not respond to the same price signals. During a scarcity event, prices rise past $80/MWh and the aluminum smelter curtails, as it always has. But the system is still short, because data center load that was not there five years ago is still drawing power and will not curtail at any price the market can reach. Prices keep rising until either new supply dispatches or enough additional traditional industrial load has curtailed even more.

For these consumers, the CHR framework provides something no demand response announcement can: a quantified measure of the asymmetry between their price tolerance and the price tolerance of the demand class that now shares their grid. That measurement is the foundation for understanding their cost exposure, evaluating procurement strategies, and eventually hedging the risk.

The Floor Problem: What Demand Response Cannot Reach

The public conversation about data center electricity demand focuses on peaks: scarcity events, extreme pricing, summer heat waves, polar vortices. Demand response is positioned as the answer to these peaks. And for a narrow set of hours, from a narrow set of operators, it provides a real, if modest, contribution. But the structural impact of inelastic data center load also impacts the floor price.

A data center is not a peaky load, it’s a pretty flat consumption. So off-peak prices rise because baseload demand is permanently higher. Shoulder-hour prices rise because the supply cushion that existed during moderate-demand periods has been absorbed. On-peak prices rise because the load that used to curtail and provide relief is now a smaller fraction of total demand, meaning the system must dispatch more expensive marginal generation.

Consider the math. A hyperscaler adds a 500 MW data center to a grid and commits 50 MW to demand response. The grid now has 450 MW of new load that will not curtail at any plausible wholesale price, running every hour of every day. Even during the hours when the 50 MW flexes, the other 450 MW has raised the structural clearing price above pre-arrival baselines.

For the industrial consumer, this means higher average annual electricity costs across all hours, with a modestly lower ceiling during the specific windows when the hyperscaler flexes. Net effect: a permanent increase in the cost floor with a marginal reduction in the worst peaks.

And this accounts only for the hyperscaler segment that can offer demand response at all.

What Needs to Be Measured

Demand response announcements tell market participants how many megawatts of flexibility have been committed. What is missing is the measurement of the load that will not flex, its price tolerance characteristics, and its impact on market clearing across all hours.

The CHR framework provides this measurement by answering four questions that demand response data alone cannot:

First, what is the aggregate inelastic load? In any given market zone, how many megawatts of data center demand are contractually or operationally incapable of curtailment at any plausible price? This is the SLA-locked colocation capacity plus the high-CHR hyperscaler workloads that will never participate in demand response.

Second, what is the price tolerance profile of that inelastic load? A single “data center demand” figure is analytically useless without knowing the CHR distribution. A market with 2 GW of data center load concentrated in commodity inference (CHR ~$800/MWh) behaves very differently from one with 2 GW concentrated in frontier inference (CHR ~$53,000/MWh). The higher the CHR of the inelastic block, the deeper the price inelasticity and the higher the structural floor.

Third, what is the ratio of flexible to inelastic load? If a market has 5 GW of total data center demand and 200 MW of committed demand response, the flexibility ratio is 4%. That number, expressed clearly and tracked over time, tells planners and policymakers whether demand response is a material resource or a narrative device. The CHR framework produces this ratio by classifying load by tolerance tier.

Fourth, what does the inelastic load do to clearing prices across all hours? Not just during scarcity events, but during the 8,000+ hours per year when the market is operating normally. The floor price effect is where the economic impact accumulates for industrial consumers, utilities managing rate cases, and policymakers evaluating cost allocation.

Flexibility at the Surface, Structure Underneath

Data center demand response is a real resource that will play a real role in grid management. It should be developed, valued, and incorporated into planning processes. Nothing in this analysis disputes that.

The flexibility layer and the structural layer need to be measured together, in a common framework, using units that allow direct comparison with the price tolerance of every other participant in the market.

The Compute Heat Rate provides that framework. It does not predict whether demand response will succeed or fail. It measures the depth of the ocean underneath the thin layer of flexibility floating on top.

More to come.

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