Steel For Fuel N° 056 of 56 21 Sep 2026
For AI, energy is nothing, and energy is everything (reprise)
by Andy Lubershane, Partner and Head of Research, Energy Impact Partners
In this note
The question
How much could an AI data center pay for electricity before the computing stops being worth doing?
The answer
Far more than the power industry is used to thinking about. A framework called the compute heat rate puts the current break-even ceiling for frontier-model inference at 142 times what American industrial consumers pay. It measures what a workload can tolerate rather than what anyone will be charged, and its author says the inputs are moving fast.
03The argument
The post is Lubershane revisiting his own two-year-old piece, and the quoted passages are the position he is building on rather than the one he holds now. In 2024 he argued that energy was a trivial share of the cost of computing, so AI developers ought to pay a substantial premium to secure power. Two things have changed. Inference, the workloads customers pay for directly, now drives demand, and there is far more public data on what it earns. Separately, the public has started watching what data centers do to electricity prices, and he quotes his own post from a few weeks earlier: an individual project probably is not raising your bill, because utilities make projects pay more than their incremental cost, but data centers collectively are putting systemic pressure on infrastructure costs.
That leaves the question he had previously declined to answer. The compute heat rate is an outside attempt at it, built by Hans Royal and collaborators, and the bulk of the post is Royal’s own explanation, invited by Lubershane and run under Royal’s headings. The name borrows from a power plant’s heat rate, the efficiency of turning fuel into electricity; this one chains electricity into floating-point operations, operations into tokens, and tokens into dollars, to give the most a facility could pay for power and still break even. Royal splits it in two, and the split is the useful part. The long-run version carries the full cost stack and a required return, so it governs whether to build. The dispatch version treats capital as sunk and governs whether to curtail a facility already running. A developer and an operator across the road from each other face different thresholds.
What keeps the headline number from being a prediction is that its two main inputs pull against each other: what matters is price per token multiplied by tokens per megawatt-hour, and the models that charge most per token are the ones that burn most electricity making each one. Token prices have already fallen substantially, and Royal says a continued trend would pull the metric down with them.
04What you need to know first
- Heat rate
- For a power plant, how efficiently it turns the energy in its fuel into electricity. The compute heat rate borrows the name for a conversion running the other way, from electricity into saleable output.
- Inference
- Running a trained model to answer requests, as distinct from training it. It is the part customers pay for directly, which is what makes revenue per unit of electricity observable.
- Token
- The unit AI providers meter and bill by. Its public price is what the framework uses to value a workload’s output.
05Details worth keeping
- Roughly two thirds of the post is written by Royal, ending where Lubershane resumes at “Back to Andy:”.
- Lubershane says the question arose when he and his partner Shayle Kann discussed it on Kann’s podcast, and that Royal wrote to him afterwards.
- The formulas, the token price history and the third-quarter 2026 index are all published as images. The index value appears only in that figure, so this note cannot carry it.
- Token throughput is estimated from MLPerf scenarios rather than measured, and self-hosted inference is proxied by GPU rental rates.
06Claims worth citing
All figures as stated on 2026-09-21. The central figure is a third party’s calculation reported by Lubershane, and it is a threshold rather than a price anyone pays.
- The compute heat rate for frontier labs selling inference is currently $12,790 per megawatt-hour. Royal and collaborators, cited by Lubershane
- Power prices for industrial consumers in the United States averaged about $90 per megawatt-hour. Lubershane
- Data centers could therefore theoretically pay up to 142 times what industrial facilities currently pay. Lubershane’s arithmetic on the two figures above
- Larger models can consume three or four times the electricity per token, so a frontier model may post only a modest compute heat rate while a small model priced at pennies posts a high dispatch figure. Royal
07Where it’s contested
Nobody contests anything here; there is no second voice arguing. What the post does carry is an unusual amount of stated uncertainty, nearly all of it volunteered by the framework’s own author.
- The framework’s author is an interested party. Royal built the metric and publishes its index, and describes it here in his own words. The caveats below are his, to his credit, but nobody independent tests the method.
- The revenue input may not be real. Published API pricing is not necessarily what enterprises pay, since they negotiate bilaterally.
- The headline number is explicitly not a forecast. It measures what a workload can tolerate, not what electricity will cost, and varies dramatically across model tiers; the blended index rests on weightings Royal calls estimates.
- It may already be falling. With token prices declining, the $12,790 is a reading at one moment rather than a stable ceiling.
- The prompting question goes unanswered. It was whether data centers could offset more of other ratepayers’ costs. The post establishes what they could tolerate paying and never returns to whether they will.