The AI Footprint ← Back to the dashboard

44 of 178 AI models cannot be assessed at all.

Not because the data is hard to find, but because their makers have never published it. This is not a gap in our figures so much as a gap in theirs.

Data built 2026-09-01

Why a model can be unassessable

Energy per answer is driven by how much of a model fires for each word it writes. Without a parameter count, published by the maker or derivable from what they have said, there is nothing to calculate from. No AI company publishes per-answer energy directly, so a size is the only route anyone outside has.

44 models here publish no size at any source. A further set publish so little that the plausible answers sit about 30× apart — those are shown as wide brackets and barred from every headline claim on this site, rather than being quietly guessed at.

Share of each line-up that can be assessed

xAI
0% · 8 models
Moonshot
40% · 5 models
Anthropic
48% · 21 models
Alibaba
64% · 33 models
Google DeepMind
65% · 20 models
Z.ai (Zhipu AI)
67% · 6 models
OpenAI
77% · 31 models
Meta AI
83% · 12 models
DeepSeek
91% · 11 models
Mistral AI
93% · 14 models
enough published to assess partly published, wide bracket only nothing published, cannot assess

Makers with fewer than 3 tracked models are left out — a percentage of one or two says more about the sample than the company.

The most capable models nobody can assess

These are not obscure releases. They are among the most capable models ever published, and no one outside the company that made them can say what a single answer costs:

Why this matters

Every other industry that consumes energy at scale is expected to say how much. A power rating on an appliance, a fuel figure on a car, a rating on a building. AI has grown to a scale where its consumption is a public question, and the basic number needed to answer it is treated as commercially sensitive.

We could have closed the gap by guessing. Assuming one architecture over another moves the answer roughly thirty-fold, so a guess would have produced a number that looked precise and meant nothing. Leaving the gap visible is the more honest result — and it is the finding, not a footnote to it.

Share this

The full comparison of 134 models · how these numbers are made · data built 2026-09-01