How we estimate what companies won't tell us

ecogpt.coPublished Oct 30, 2026 · 2 min read

161 of the 210 key figures in the index are estimates or derivations, because 27 of 30 companies publish nothing per prompt and many publish nothing at all. Every estimate shows its working. Here are the rules, the factors and the uncertainty.

Energy per prompt is anchored on the closest measured analogue and scaled by architecture; uncertainty about ±3×.
Water is energy × fleet WUE, on-site boundary; carbon is energy × the serving grid's intensity, location-based.
Annual emissions for non-reporters are fleet power × hours × 80% × grid, compute only.
An estimate never raises a score and is replaced by a disclosed figure on receipt.

Why estimate at all

Because a blank is the one number nobody can argue with. Leaving a dash where OpenAI's carbon per prompt should be would let the absence pass unremarked. A labelled estimate with its working printed invites the company to replace it, and lets readers judge the reasoning. The index's rule is that estimates never improve a score: the efficiency pillar rewards disclosure, so an estimated figure earns the same confidence grade as none.

The rules

Energy per prompt. Start from the closest measured analogue: Epoch's 0.3 Wh for a dense frontier model like GPT-4o; the 0.042 Wh median for sub-10-billion-parameter models from Luccioni et al.; 1.6 to 3.6 Wh for a 1024-pixel image. Scale by architecture: a mixture of experts with 17 billion active parameters gets about a third of the dense figure; a retrieval-augmented search answer gets two to three times a chat turn. State the uncertainty, typically ±3×.

Water per prompt. Energy × the fleet's water-use effectiveness, on-site boundary, so the result is comparable with Google's and OpenAI's. Where the company runs on someone else's cloud, inherit that cloud's WUE (AWS 0.15, Azure 0.30, Google 1.15). Print the full-scope figure, adding 1.8 L/kWh for electricity generation, in the note.

Carbon per prompt. Energy × the average intensity of the grid where the prompts are served, location-based: 390 g/kWh in the US, 580 in China, 55 in France, 500 on gas turbines. Certificates are not netted off; the note says so.

Annual emissions for non-reporters. Fleet power in MW × 8,760 hours × 80% utilisation × grid intensity. Compute only, not comparable with a corporate inventory, and labelled as such.

Fleet power and chips. Disclosed electricity ÷ hours where it exists; otherwise accelerators × 1.3 kW all-in, the ratio observed at Colossus; otherwise an order-of-magnitude estimate from announced capacity or capital spending, with ±2× or worse stated.

The factors

Every constant lives in one file and is listed with its source on the methodology page: the grid intensities, the WUEs, the kilowatts per chip, the 514 kg of embodied carbon per accelerator slot, the 17 kg of server, the 27 tonnes of copper per megawatt, the 2.6 MW per generator. Change one and every derived figure on the site changes with it.

What the labels mean

"Company figure" is the company's own. "Measured independently" is a third party's. "ecogpt.co estimate" is ours, with the method under the number. "Derived" is arithmetic on one of the first two. "Not disclosed" appears only where no defensible estimate exists, which after this edition is almost nowhere.