• Metric defined as maximum electricity price at which AI load remains economic
• PJM is considering CHR-informed demand response for flexible data center loads
• Frontier AI workloads can justify electricity prices far above conventional market levels

Hans Royal did not set out to create a new metric for the AI power boom. He was trying to understand whether the boom itself was real.

Royal, an industry veteran, spent most of his career in energy, especially renewables, but took a sabbatical last year and enrolled in an in-person course at MIT to get a firmer grip on artificial intelligence. He said he wanted to know whether AI was “legit,” whether it would matter for his life and family, and whether it might reshape the energy business. Out of that came a simpler question: what, exactly, is a megawatt-hour worth to an AI model?

That question led Royal to what he calls the Compute Heat Rate, or CHR, a metric meant to do for AI demand what the gas heat rate did for generators. In his methodology paper, Royal defines CHR as the maximum electricity price at which a given AI workload remains economic after accounting for non-electricity costs and a required margin. In plain English, it is the price ceiling at which an AI operator would rationally curtail. Royal frames it as the demand-side analogue to the gas heat rate: where the gas heat rate converts fuel cost into generation cost, CHR converts compute revenue into maximum tolerable electricity cost.

“It has helped me understand the marketplace a little bit better, at least through my specific lens, which is the energy-to-intelligence intersection,” he said in an interview, which he granted independently of his role at Schneider Electric.

His baseline results underscore why the concept has drawn attention. Royal’s framework puts frontier inference at roughly USD 53,650/MWh in the first quarter of 2026, mid-tier inference at about USD 8,120/MWh and a blended average across AI workload types at roughly USD 6,350/MWh, or about 127 times the conventional gas heat rate benchmark of around USD 50/MWh.

Royal compared CHR to a thermometer: first a way to measure something previously fuzzy, then a way to price, hedge and manage the risk around it. He said industrial users such as steel mills might eventually use it to think about hedging electricity-price exposure in markets where AI load is tightening supply. He also said investors could use it to judge whether forward power curves understate the pressure that data centers may exert on prices.

“You can hedge things that you can measure – everyone else besides the data centers might want to actually hedge their exposure, and if you have a good measurement for the risk, then you can start hedging it, and there’s different counterparties for that,” he said.

He also used the concept to frame the value in NextEra’s merger with Dominion Energy: “When NextEra looks at Dominion, they are seeing the highest-CHR-impacted node in North America: a grid location where the demand class generates USD 6,350 in revenue per MWh of electricity consumed, where that demand class is growing faster than anywhere else in the country, and where capacity prices have already repriced by an order of magnitude,” he wrote, referring to capacity auction prices that went from USD 28.92/MW-day to USD 333.44/MW-day across just three auction cycles, reaching the FERC-imposed cap.

Market impact

The idea is already beginning to seep into market design. PJM’s May white paper on reliability explicitly cited CHR and used Royal’s workload tiers to argue that current price caps may be too low to unlock economically rational demand response from sophisticated data centers.

PJM said that at prices closer to USD 10,000/MWh, lower-value AI workloads could become economic to curtail, while high-end frontier workloads would keep running. The paper also pointed to a new kind of graduated response: data centers trimming 10%, 20% or 30% of load in seconds through software-based controls, batteries and workload shifting rather than shutting off entirely.

“This is precisely the graduated, price-elastic demand response the grid needs: not a binary shutdown, but a tiered response where the most flexible compute responds first, in proportion to price,” PJM wrote in the paper.

“PJM is saying basically, ‘Hey, if we can measure the data center’s actual CHR levels across various workloads, we could actually create structures that allow for demand response at much higher levels than current demand response programs,’” Royal said.

His latest Substack post suggests those thresholds are not standing still. Writing after Anthropic released Claude Fable 5 on June 10, Royal calculated a CHR of roughly USD 263,000/MWh for Fable 5 inference using the same reference configuration he used for GPT-5.5, allowing an apples-to-apples comparison.

That placed Fable 5 above OpenAI‘s GPT-5.5 standard at USD 156,000/MWh but below GPT-5.5 Pro at USD 955,000/MWh. Royal argued that two frontier-model price doublings in seven weeks showed the metric moving higher, not lower, despite efficiency gains.

The combination of more efficient AI chips and more efficient frontier models charging higher prices, in other words, continues to drive the CHR higher. Can it ever come down?

“If eventually API prices come down, CHR would go down. If there’s some physical limit, and there probably is, in the chip design and the algorithms where it flattens out in terms of how many tokens they can get out per chip or equivalent megawatt hour of throughput, then you can have a race to the bottom in terms of cost and then maybe CHR goes down over time,” he said.

“So those are the examples of where that could happen,” he added. “It just hasn’t happened yet.”