Every compute-to-energy conversion in circulation multiplies nameplate TDP by an assumed PUE. A 700-watt H100, a 1.3 power-usage factor, 0.91 kilowatt-hours per GPU-hour, done. That number underpins data-center demand forecasts, interconnection studies, compute index methodologies — and, until this piece, four dashboards on this site.

In April, the National Laboratory of the Rockies measured it instead. Vercellino et al. instrumented an H100 cluster at 0.1-second resolution across MLCommons training and fine-tuning benchmarks and vLLM inference, then scaled the measured profiles to whole-facility demand with a discrete-event model. Crucially, they published everything — the raw traces, the simulated facility series, and the measurement tooling — under CC-BY-4.0. You can re-run it, extend it, and publish derived work with attribution. That is rare enough to be the story on its own.

I pulled the dataset and rebuilt this site’s compute energy layer on it. Three things came out that change how the numbers should be quoted.

1. Real workloads draw 12–21% less than nameplate — and that is not the gpu-burn gap

The paper’s headline device figure is 668.2 W ± 1.4 against a 700 W TDP under gpu-burn — a 4.5% gap. That is the number most readers will take away, and for an energy-cost model it is the wrong one.

gpu-burn is a synthetic stress ceiling: what a GPU draws pinned at maximum. A real workload is not pinned. It checkpoints, evaluates, loads data, and dips between phases. Computing sustained draw across the full busy period of each published trace gives duty factors of 0.795 to 0.880 — meaning nameplate overstates device-level energy by 12% to 21%, averaging about 16%. Stable Diffusion training sits at the bottom (0.798) with a p10–p90 band of 0.777–0.976, which is the numerical signature of the 12→48 kW oscillation the authors describe during iterative denoising at evaluation.

Both numbers are correct; they answer different questions. Ceiling sizing wants gpu-burn. Energy cost — dollars per GPU-hour, and the denominator of any energy-normalized index — wants the duty factor. Conflating them is a 4.5% error where a 16% one belongs.

2. At facility scale the gap widens, and splits in two

The simulated 10 MW colocation facility peaks at 73.2% of rated IT power even at high occupancy; the 1 MW inference facility at 80.3%. Those are the derates worth quoting — but only for peak questions.

For energy questions the right figure is the capacity factor, which at 80% target utilization is 59.1% (colocation) and 66.3%(inference). And because the facility simulation already embeds the workload power profiles, the peak derate already contains the duty-factor effect — multiplying them together double-counts. The calculator below asks which question you’re answering and branches accordingly, because in practice this is where announced-gigawatt arithmetic goes wrong.

One corollary falls out of the load duration curve: colocation at 40% utilization spends 673 hours a year — 7.7% of the time — above 95% of its own peak. Everything upstream of the meter is sized for a condition that holds one hour in thirteen.

3. We were using nameplate too

This site’s energy-normalized compute analysis states it in its own footer: “GPU TDP figures are nameplate.” The Compute × Power workbench uses the same convention. Both inherit the error described above. The second tab of the calculator recomputes the affected figure — the energy content of a GPU-hour, priced against the 374,550 hourly ERCOT and PJM settlements behind the hourly power piece — under both conventions, so the correction is visible rather than quietly patched.

What this means for an energy-normalized index

ICE and NATIVX have announced compute futures on the COIL Index, which tracks tokenized, energy-normalized compute — explicitly designed to “strip out the noise introduced by regional power cost disparities.” Energy normalization makes the index structurally a ratio: compute price divided by energy content.

The numerator is observable from transactions. The denominator is a modeling assumption — and this dataset is the first public evidence of how the obvious assumptions behave. A nameplate-derived denominator overstates energy content by 12–21% at the device level and more at facility scale; a fixed kWh-per-GPU-hour ignores that measured energy per unit of work varies by workload type, by node count, and by request rate. The paper’s inference finding sharpens this: power flattens while throughput keeps rising, so energy per token falls as load rises. Two facilities running identical hardware at different request rates have different energy content per unit of compute delivered.

The conclusion is not that normalization is wrong. It is that an energy-normalized index does not eliminate assumptions — it relocates them from power geography into workload mix. COIL carries an embedded workload-mix basis that stays invisible unless the constituent basket’s utilization and job composition are disclosed. That is a fair question to put to any benchmark administrator, and it maps onto IOSCO Principles 6 and 7 (benchmark design and data sufficiency) and Principle 11 (methodology content): does a denominator derived from nameplate ratings adequately reflect the underlying interest?

I ask that having just corrected the same error in my own published work. The methodology document is not out yet; if COIL already uses measured or transaction-derived energy content, this piece becomes corroboration rather than critique, and the underlying point still holds — the denominator deserves to be published, whatever it is.

The caveat that matters most

This is H100 silicon only. No H200, no B200, no GB200 NVL, no liquid-cooled or HVDC rack architecture. Blackwell-class systems will have materially different power-per-unit-work characteristics, and the duty factors above should not be extrapolated to them. The facility simulations also exclude cooling and non-IT load, and their utilization shapes are borrowed from Shanghai AI Lab and Azure traces rather than observed at a commercial hyperscale site — which makes them illustrative structure, not forecast. The authors flag all of this themselves.

None of it undermines the usable core: measured node-level power under standardized, reproducible benchmarks, published openly. It is the best public calibration point that exists for the H100 generation, and the correct response to its limits is to extend the measurement — which, given the license and the open tooling, anyone can now do.

Power data derived from Vercellino, Willard, Campos, da Silva Pereira, Hull, Selensky & Mueller, “Measurement of Generative AI Workload Power Profiles for Whole-Facility Data Center Infrastructure Planning”, arXiv:2604.07345 (CC-BY-4.0); dataset DOI 10.7799/3025227, National Laboratory of the Rockies. Duty factors, bands, capacity factors, and load-duration statistics computed from the published traces and simulated facility series. Hourly power prices from ERCOT MIS and PJM Data Miner 2, July 2023 – July 2026. GPU rental marks are illustrative. Research and education only — not investment advice.