The Compute Heat Rate
A Preview of What's Coming for Electricity Markets
I’ve spent the last several months building an analytical framework that I think changes how we should be thinking about wholesale electricity markets. This is a preview. The full research is coming. But the core idea is simple enough to share now, and I think it matters enough that it shouldn’t wait.
Here’s the setup.
In natural gas power generation, there’s a concept called the “heat rate” that tells you how efficiently a gas plant converts fuel into electricity. It’s one of the most important numbers in energy economics because it connects gas prices to power prices. If you know the gas price and the heat rate, you can estimate the marginal cost of generation. Simple, foundational, used by everyone in the industry.
I’ve been working on an equivalent metric for AI computing. I’m calling it the Compute Heat Rate™, or CHR1.
Just like the gas heat rate connects fuel prices to power prices, the Compute Heat Rate connects AI economics to electricity demand behavior. It answers a specific question: What is the maximum price a data center operator can rationally pay for a megawatt-hour of electricity before the computation running on that MWh becomes uneconomic?
I think the CHR deserves to be tracked over time the same way the industry tracks gas heat rates. Even if my specific numbers evolve as the data improves, the metric itself matters. It tells you the price ceiling that this new category of demand is willing to pay, and that ceiling is what sets the upper bound on where wholesale electricity prices can go in supply-constrained regions.
The answer, when you actually run the CHR calculation across different AI workload types, is startling.
Traditional industrial electricity consumers become uneconomic at surprisingly low price points. According to the Aluminum Association, smelters need electricity at or below $40/MWh on long-term contracts to remain viable. The U.S. went from 33 operating smelters in 1980 to just 4 today, largely because electricity got too expensive. Steel electric arc furnaces start shifting production around $60 to $80/MWh. Broader industrial demand response programs in PJM and ISO-NE typically activate when wholesale prices cross $75 to $100/MWh. These thresholds have acted as a natural brake on electricity prices for decades. When prices spike, industrial load drops off, and prices come back down. The market has relied on this self-correcting mechanism for so long that most people don’t even think about it.
AI data centers don’t work this way.
When you calculate the full-stack economics of running AI workloads (GPU costs, facility overhead, cooling, networking, and the revenue generated per MWh consumed) the Compute Heat Rate is not $40 or $100. It’s not even in the same universe. Depending on the workload type, the CHR ranges from roughly $250/MWh on the low end to well over $3,000/MWh on the high end.
Again, the CHR for the lowest-margin AI workloads is 3 to 6x higher than what triggers demand response from heavy industry and for the highest-margin workloads, the CHR is 30 to 60x higher (!!).
This isn’t speculation, this is derived from publicly available GPU pricing, cloud compute rates, API pricing data, and reported infrastructure costs. The math is pretty straightforward once you set up the framework.
So what does this actually mean for electricity markets?
It means we’re adding a massive new source of electricity demand that doesn’t respond to price signals the way every other large load in history has responded. When wholesale prices spike, AI load stays on. The economics support it, and the competitive dynamics of the AI industry don’t allow for voluntary downtime. The Compute Heat Rate tells us exactly why: the revenue these workloads generate per MWh consumed is so high that even extreme electricity prices don’t make the math negative.
The practical effect: the historical demand-side brake on electricity prices is being removed. Not weakened. Removed. In regions where data center load reaches critical mass relative to available supply, prices can rise to levels that would previously have triggered enough demand destruction to bring them back down. But now they won’t come back down, because the CHR tells us the new marginal buyer can afford to keep bidding at prices that would shut down every aluminum smelter and steel mill in the country.
The floor for long-run electricity prices in these regions is set by the Cost of New Entry, or CONE, which is what it costs to build new generation capacity. That’s roughly $80 to $130/MWh depending on technology and location. The ceiling is set by the Compute Heat Rate. And the gap between those two numbers is enormous.
Why hasn’t this shown up in prices yet?
Two reasons. First, the installed base of AI data center capacity hasn’t crossed the critical threshold in most regions. There’s a massive pipeline of projects in development (estimates range from 65 to 175 GW depending on who’s counting and how you define “committed”) but the majority are still in permitting, financing, or early construction. The physical infrastructure takes time to build.
Second, forward price curves in electricity markets are notoriously bad at pricing structural demand shifts before they physically manifest. The curves reflect historical patterns with modest adjustments. They don’t know how to price a step-change in demand that hasn’t fully arrived yet. This is the same dynamic that played out with California’s duck curve. It was theoretically predictable for years before solar penetration made it visible in the data. The Compute Heat Rate could serve the same early warning function for AI-driven repricing that the duck curve eventually served for solar. The difference is that this time we can see it coming and measure it before the damage shows up in the price data.
The bottom line for corporate energy buyers.
If you’re a large electricity consumer and you’re benchmarking your PPA economics against consensus forward curves, you may be looking at a model that doesn’t account for the Compute Heat Rate. And if your model doesn’t account for the CHR, it doesn’t account for the most significant structural shift in electricity demand in decades. Current PPA pricing in most markets reflects a world where AI demand growth is either ignored or dramatically underestimated.
I’m not saying the sky is falling. I’m saying the risk is asymmetric. If this thesis is wrong and AI demand stalls, a PPA at today’s prices is still a competitive energy cost with sustainability value. If this thesis is right, an unhedged position could mean tens of millions of dollars per year in incremental costs for a large industrial consumer.
The cost of hedging is modest. The cost of being exposed is not.
I’ve got a lot more detail behind this. A full research report with CHR calculations by workload type, a scenario modeling framework, settlement hub-level analysis, and a tool that lets you stress-test your own assumptions about where the Compute Heat Rate takes wholesale prices over the next decade. I’ll be sharing more in the coming weeks. This is just the preview.
If you work in energy procurement, corporate sustainability, or energy finance, I’d genuinely like to hear your reaction. Does this framework resonate? What am I missing? What questions does it raise?
More soon.
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.
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