GPT-5.5 Just Dropped and the Compute Heat Rate Just Jumped.
Every model release makes AI electricity demand more price-inelastic, not less. Here’s the math.
OpenAI released GPT-5.5 today, which I think is just about six weeks after GPT-5.4. It’s now twice the API price per token. So, the release cycle is accelerating, and each time a new model is released, the capabilities and benchmarks increase (obviously). As such, the price per token keeps rising and every one of these variables pushes in the same direction for electricity markets.
So let’s translate the model announcement into grid economics, using the Compute Heat Rate metric.
---
The Compute Heat Rate™ (CHR)1 measures the maximum electricity price an AI workload can sustain before the operator would rationally curtail. The formula is straightforward: take the revenue generated per MWh of electricity consumed, subtract non-electricity operating costs, and apply a margin requirement. What falls out is the hypothetical price ceiling for that demand class. In other words, the data center’s hypothetical willingness to pay for electricity.
GPT-5.5 standard tier is priced at $30 per million output tokens. Running on a current-generation NVIDIA B200 GPU at roughly 2,500 output tokens per second (conservative, 50% utilization), that GPU generates about $270 per hour in revenue. It draws about 1 kilowatt. Apply a facility-level PUE of 1.3, subtract non-electricity costs (GPU amortization, cooling, networking, maintenance), apply a 30% margin requirement, and you get a Compute Heat Rate for GPT-5.5 frontier inference of approximately….drumroll….
$156,000 per MWh!
That is roughly 3,100 times the gas heat rate.
For context, the blended CHR across all AI workload types published in my SSRN paper is approximately $6,350/MWh (127x the gas heat rate). The frontier inference tier has always been the high end. But $156,000/MWh is mind-bending new ceiling.
And then there’s GPT-5.5 Pro, priced at $180 per million output tokens. Same GPU, same power draw, same electrons. Six times the revenue. The CHR for that tier approaches $955,000/MWh. Nearly a million dollars per megawatt-hour before the operator would consider turning it off. I’m not even going to use that number it’s so absurd.
---
The instinct is to dismiss these numbers as theoretical extremes. Nobody is actually paying $156,000 for a megawatt-hour of electricity, and they never will. THIS IS TRUE and something people routinely misunderstand about CHR. It misses the point entirely.
CHR doesn’t predict the specific price of electricity. It measures the depth of demand-side price inelasticity. It answers the question: if electricity prices spike, who curtails first?
A steel mill operating at thin margins curtails at $80 to $160/MWh. A data center running GPT-5.5 inference under customer SLAs would keep running at $500/MWh, $1,000/MWh, $5,000/MWh, and never blink. The electricity cost is a rounding error in the value chain. At $500/MWh, electricity represents less than 0.3% of the revenue that GPU is generating.
---
Here’s the part that should concern energy planners: the CHR is going up with every model generation, not down.
A standard objection is that AI is getting more efficient. Per-token costs have fallen roughly 1,000x since GPT-3. Shouldn’t that mean less electricity demand, lower price tolerance, a softening of the CHR thesis?
No- it’s actually the exact opposite. This is the Efficiency Trap I’ve written about previously, and GPT-5.5 is the latest proof point.
Every time the model gets smarter, not only does demand increase for tokens (Jevon’s paradox), but the price per token also rises, which means the value of each MWh passing through a GPU rises, too.
Electrons just got more expensive.
---
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
Hi Hans,
This is all very interesting. When calculating the CHR, are you including any revenues for the input tokens? In the base CHR methodology you note pricing for both input and output, but in this example it appears to only include the 2500 tokens/second of output (and why 50% util as an aside?). Including input token revenue would seemingly push the CHR even higher.
And when you try to account for the impact on jobs, taxes, deficits and borrowing implications alongside household consumer surplus reduction the whole equation points in one direction for society 😳 productivity gains are dwarfed by the micro and macro disbenefits