Planned research area · No standalone observatory published
What is
visible?
Independently evaluating AI’s environmental impact remains genuinely difficult. This area is intended to bring environmental disclosure analysis into one place.
Conceptual illustration. Not measured data.
What this area
covers.
This area is an honest accounting of what can be found in public records and what cannot: which providers publish measured energy figures, which publish statements without a method, and which publish nothing, and through what channel and when.
It follows one rule: report what a document says, never characterize what it means. A gap is recorded as a gap, not as a verdict.
No standalone analysis is published here yet. The findings below are from Candor's review in October 2026. Each is a statement about what was found, not a claim about what exists: a search is not proof that nothing has been published.
What is publicly
reported.
- Google has published a measurement of energy for its median Gemini Apps text prompt: 0.24 Wh, from production instrumentation of its own serving. Elsworth et al., arXiv:2508.15734
- OpenAI's CEO has stated that the average ChatGPT query uses about 0.34 Wh, without a stated model, prompt length, boundary or method. blog.samaltman.com
- Meta publishes training energy and emissions for each Llama model in its model cards, and nothing about inference. Llama 3.1 model card
- Mistral has published a lifecycle analysis: 1.14 g CO₂e per 400-token response, which is a carbon figure that includes manufacturing, not an electricity figure. Mistral AI
- For Anthropic, Microsoft, xAI, Amazon and Cohere, Candor found no per-prompt energy figure from the company.
What remains
unknown.
- Which facility serves a given AI request, and what share of any facility's load a provider's workloads represent.
- Measured energy per prompt or token for the providers listed above, other than Google.
- Whether published per-prompt estimates are too high or too low for current models.
- Facility-level electricity consumption and power sources beyond what public records show.
Intended scope,
not published findings.
- A disclosure matrix: provider, type of information, channel (report, filing, blog, paper), and date.
- The documented-events pipeline: regulatory filings and orders tied to a specific facility. It starts empty, and nothing publishes until it clears the same primary-source standard as everything else.
- A running record of changes, so a new disclosure is dated and a correction is visible.
Questions this area asks
- Which kinds of energy information are published, by whom, and in what form?
- Where a figure is published without a method, what can and cannot be concluded from it?
- How do disclosures change over time?