🎙 THIS WEEK ON THE ENERGY.MEDIA PODCAST
Hans Royal: The $6,200 Number That Breaks the Forward Curves
I sat down with Hans Royal, creator of the Compute Heat Rate framework, to break down why power markets are about to get the biggest mark-to-market correction in their history. PJM cited his work by name in the May 6 capacity reform white paper. We get into all of it.
▶ Watch the full episode on YouTube | 🎧 Listen on Spotify Podcasts⚡
⚡THE BIG STORY
Why I Brought Hans Royal on the Show
Every power trader I talk to is staring at the same disconnect and shrugging. Vistra, NRG, and Constellation are trading at multiples that only make sense if power prices are headed up. Gas turbines are sold out through 2029. Hyperscalers are filing interconnect requests for hundreds of gigawatts. And yet forward curves in ERCOT and PJM look soft. Year-over-year, traders are calling the market “depressed.” Spot gas just hit a 16-month low. Nobody in the room has a clean answer for why the curves and the equity tape are telling opposite stories.
That’s the conversation I wanted to have on the podcast. So I called Hans Royal.
Hans is the creator of the Compute Heat Rate — a framework I’ve been quietly tracking for months that I think is the single most important analytical tool that’s emerged in the AI-power conversation. PJM cited his work by name in the May 6 capacity market reform white paper. They built the entire Hyperscale Paradigm Shift section of that document around it. If you read one piece of analysis on power markets this year, it should be his.
The full episode is up now on the Energy.Media podcast. Below I’ll give you the operator’s read on what came out of the conversation. But the show is where the real value is — Hans walks through the methodology, the workload tiers, and the math in a way that takes 45 minutes to do justice. This newsletter is the appetizer. Go watch the episode.
THE FRAMEWORK: VOTING MACHINE VS. WEIGHING MACHINE
In the short run, the market is a voting machine. In the long run, it’s a weighing machine. — Benjamin Graham, by way of Warren Buffett.
That line was written about stocks. It applies just as cleanly to power markets, and it’s the cleanest way I’ve found to explain the disconnect Hans and I unpacked on the show.
Look at what the voting machine is doing right now.
Natural Gas Intelligence reported in March that gas-fired generation averaged 149 GW for the week ended March 13 — down 12% week-over-week and the lowest weekly level since last spring. Wind generation surged 39% to 78 GW. Coal-fired output fell 14%. Total generation dropped to 443 GW, the lowest week since November. Gas flows to power plants hit 28.0 Bcf/d, the lowest level since April 2025. Spot gas prices tumbled to a 16-month low at $1.925/MMBtu. Heating degree days came in at 101 versus a 10-year average of 148 — a 32% miss to the downside.
Mild weather. Strong wind. Soft demand. Soft prices. That’s the voting machine. And every trader pricing the forward curve is feeding that data into a regression model that extrapolates the recent past forward. The curve goes down. The desk says “market is depressed.” Forward power in ERCOT trades at numbers that wouldn’t cover the financing cost of a new combined cycle plant, let alone justify breaking ground on one.
Here’s the problem: the voting machine is voting on March weather. It is not voting on the demand-side structural change that is six to thirty-six months out from energizing on the grid.
Now look at what the weighing machine is going to weigh.
This is where Hans’s work comes in, and where the conversation got sharp on the show. The Compute Heat Rate measures the maximum price an AI data center can pay for electricity before its economics break. The gas heat rate tells you the ceiling on what a combined cycle plant can pay for fuel. Compute Heat Rate flips it: it tells you the ceiling on what an AI workload can pay for power.
“That number actually keeps going up over time, not down, which is a little bit counterintuitive. What that means then is from just a raw economics perspective, an AI data center does not voluntarily want to economically curtail themselves until the price of power goes above, in some cases, the statutory limit in these markets.”
— Hans Royal, on the Energy.Media podcast
The blended average CHR across AI workloads is roughly $6,200 per MWh. That’s 127 times the going wholesale price in most U.S. markets. The conservative low end is $250 to $500 per MWh for commodity training runs. The frontier inference end — GPT-5.5 and Opus-class models — runs to $50,000 to $150,000 per MWh equivalent. Hans walks through the methodology on the show: he uses publicly available API token pricing, server CapEx, and a 30% margin assumption. Conservative inputs, eye-watering outputs.
And the kicker: as Hans keeps updating his research, the number keeps going up, not down. The new Nvidia Vera Rubin chip is reportedly 5 to 10x more efficient than its predecessor. Counterintuitively, that raises the value of an electron rather than reducing it. More tokens per kilowatt-hour means each kilowatt-hour is worth more. Jevons Paradox in the data center.
This is the demand-side variable that none of the forward curve methodologies are pricing in. When prices spike to $150 or $300, an aluminum smelter goes home. A steel plant runs shift one and skips shift two. A data center running Claude Opus inference doesn’t even notice. The economic curtailment break that has cleared scarcity events for fifty years doesn’t exist for this load class. And we’re about to add a hundred-plus gigawatts of it.
The voting machine and the weighing machine are about to collide.
Forward curves are heavily driven by short-term weather and trader sentiment. They are exquisitely sensitive to the last storage report, the last temperature outlook, the last wind print. They are almost entirely insensitive to the structural demand pipeline that takes 18 to 36 months to show up. That mismatch is the trade.
In the short run, mild March weather will keep gas burns down and the curves soft. In the long run, ERCOT and PJM are going to add load that doesn’t curtail at any price the curve currently contemplates. Both can be true. Both are true. The short-run signal is real. The long-run signal is also real. The market is voting today. It will weigh tomorrow.
Power producers who own dispatchable generation are about to get re-rated. The forwards aren’t giving it to them yet. The equity markets already are. The equity tape is reading the weighing machine. The curve desk is still reading the voting machine. Eventually, one of them is wrong.
The full conversation — including how Hans built the framework, why he thinks the bubble argument is wrong, and what specific workload tiers tell you about which markets get hit first — is up on the podcast now. This newsletter is the operator’s synthesis. The episode is where the analyst lays out the work.
🎙 THIS WEEK ON THE ENERGY.MEDIA PODCAST
Hans Royal: The $6,200 Number That Breaks the Forward Curves
I sat down with Hans Royal, creator of the Compute Heat Rate framework, to break down why power markets are about to get the biggest mark-to-market correction in their history. PJM cited his work by name in the May 6 capacity reform white paper. We get into all of it.
▶ Watch the full episode on YouTube | 🎧 Listen on Spotify Podcasts⚡
💰 DEAL SHEET
PJM publishes capacity market white paper — May 6. A 70-page document from David Mills and the PJM staff laying out three structural reform paths. The paper cites Compute Heat Rate as evidence that current price caps are below what AI workloads can pay — using Hans’s framework to argue for raising the Synchronized Reserve ORDC penalty from $850/MWh to $2,100/MWh. This is a regulator validating a piece of analysis you should know.
NGI: Power burns hit yearly low. Gas-fired generation averaged 149 GW for the week ended March 13, down 12% week-over-week. Wind surged 39%. Spot gas fell 26.5 cents to a 16-month low. The voting machine doing what it does.
Anthropic reportedly hits $30B+ ARR run rate. The number is moving up roughly 10x annualized. Whatever the bubble debate is, that’s not bubble revenue — that’s revenue growth showing up in the API tokens that drive the CHR math.
GPT-5.5 priced at 2x GPT-5.4. OpenAI doubled the API cost of its frontier model on launch. Token prices going up at the frontier is the mechanical driver of CHR going up. We get into this on the show — every model release that prices higher pushes the ceiling on what the data center running it can pay for power.
Nvidia Vera Rubin announcement. Reportedly 5–10x more token output per chip. Hans’s counterintuitive read on the show: this raises CHR, not lowers it. More tokens per electron means electrons are worth more.
📊 DATA POINT OF THE WEEK
$1.9 vs. $6,200.
$1.9 is where spot gas was trading in mid-March, a 16-month low driven by mild weather and strong wind. $6,200 is the blended average Compute Heat Rate — the price an AI data center can pay for electricity before its economics break. One of those numbers is the voting machine. The other is the weighing machine. The forward curves are pricing in the first and ignoring the second. Hans walks through the full methodology on the podcast — it’s also published on computeheatrate.com. The point isn’t that power prices will hit $6,200. The point is that the demand-side curtailment break the forward curve methodology depends on doesn’t exist for this load class until prices are an order of magnitude above where they currently sit.
🔥 HOT TAKE
Forward power curves are about to get the biggest mark-to-market correction in the history of competitive power markets, and the people running them don’t see it coming.
I’m not saying power prices go to $6,200. I’m saying the methodology that produces today’s forward curves is structurally incomplete. It assumes a demand curve that bends at industrial-load price levels. That assumption was correct for sixty years. It is no longer correct, and the people building the curves are using the same regression models they used in 2019, fed by the same weather and storage inputs that drove this week’s spot price to a 16-month low.
The voting machine works in the short run because the inputs are observable: weather forecasts, wind output, storage levels, daily burns. Those inputs feed cleanly into a model. The weighing machine works in the long run because supply and demand fundamentals always reassert themselves — but the inputs are messier: the share of announced data center load that actually energizes, the speed at which gas turbines arrive, the pace of new build relative to retirements, the willingness-to-pay of a load class that has never existed at this scale.
Hans’s number is conservative. He stress-tested it down to $250–$500/MWh on the low end. Even at the most conservative blended average — call it $1,500/MWh, a quarter of his published number — you are still 30x above today’s wholesale prices. There is no version of this math where the forwards are correct over a multi-year horizon. There is only the question of how violent the correction is when reality starts hitting the curves.
I want to be careful about how I say this, because I’m a power producer and my book is long this thesis. But the structural argument doesn’t depend on me being right about magnitude. It depends on three things being true: (1) AI data centers will continue to monetize compute at rates that produce CHR numbers in the thousands, (2) enough of them will actually energize on the grid to matter for marginal pricing, and (3) the forward curve methodology does not currently price in any of this. All three are true today. The first two are getting more true every quarter.
Go watch the episode. Hans makes the case better than I can summarize it here.
You may end up doing better than market price signals currently indicate.
📡 ON MY RADAR
Summer load forecasts and the first real heat event. The voting machine that drove March prices low can flip fast. The first sustained heat dome over Texas this summer is going to test how soft these curves really are. Watch ERCOT scarcity pricing in July and August.
PJM stakeholder process on the May 6 white paper. Comments and counter-proposals will start landing in June. Watch how LSEs and state regulators react to the credibility trap framing — that’s the political pressure point that determines which of the three paths actually moves.
Q2 earnings from Vistra, NRG, Constellation, and Talen. The disconnect between equity multiples and forward curves shows up most clearly in management commentary on capacity revenue and PPA pricing for new build. Listen for what they’re saying about hyperscaler procurement.
🎙 THIS WEEK ON THE ENERGY.MEDIA PODCAST
Hans Royal: The $6,200 Number That Breaks the Forward Curves
I sat down with Hans Royal, creator of the Compute Heat Rate framework, to break down why power markets are about to get the biggest mark-to-market correction in their history. PJM cited his work by name in the May 6 capacity reform white paper. We get into all of it.
▶ Watch the full episode on YouTube | 🎧 Listen on Spotify Podcasts⚡
If this analysis is useful, forward it to a colleague who’s pricing power deals. And if someone sent this to you, subscribe at powersignal.substack.com so you don’t miss next week’s issue.
