How to substantiate AI energy claims
A checklist for sustainability, legal and marketing teams who publish energy figures for AI products and infrastructure, as of 5 November 2026.
What counts as an energy claim
Any statement about how much energy a model, product or facility uses: per query, per token, per task, per year, or compared with an alternative. A claim is only as strong as the boundary and the period behind it.
The EU rule from 27 September 2026
- Directive (EU) 2024/825 applies from 27 September 2026, with no transition period (Mayer Brown, 27 October 2026).
- Generic environmental claims such as "green" or "eco" are banned unless recognized excellent environmental performance can be shown.
- Product-level climate-neutral claims based on offsets are banned.
- Claims about future performance need a verified implementation plan, checked by an independent expert.
- It covers business-to-consumer practices. Member States may extend it to business-to-business claims, and as of 27 October 2026 eight had notified no national measures.
In the United States
The FTC's Green Guides are guidance rather than regulation, and expect environmental claims to be backed by competent and reliable evidence (FTC Green Guides, 2012 revision (16 CFR Part 260)). A summary updated on 5 August 2026 reports that the Guides remain in effect while the FTC's review continues (netzerocompare.com summary, updated 5 August 2026).
A boundary checklist
- Name the energy account: work, overhead or total.
- Name the meter point: chip, board, system or facility.
- State the period and the task set.
- State utilization and batch size.
- Say whether facility overhead (PUE) is included.
- Label the figure Measured or Estimated; see labels and grades.
Measured versus modeled
Many published per-query figures are modeled rather than metered. One widely cited 2025 study estimated about 0.42 Wh for a short query to one model, using rated power, assumed utilization and a PUE multiplier, and found that changing the batch size moved per-prompt energy by about 45% (Jegham et al., 2025). A modeled figure is useful, but it must be labeled as one.
Wording that states its boundary
| Weak | Stronger |
|---|---|
| "Our AI is green." | "2.1 Wh per completed task at M3, measured over 10,000 tasks in October 2026, excluding facility overhead." |
| "50% more efficient." | "50% more passing tasks per kWh than our 2025 system, both measured at M3 on the same task set." |
| "Climate-neutral inference." | A figure with its boundary, period and emissions method, and no neutrality claim based on offsets. |
Not legal advice
This guide is general information, not legal advice. Check the current text of each rule, and take advice for your situation.
Questions
Does Directive 2024/825 apply to B2B data-center claims?
It is a consumer-protection directive, so not automatically. It may matter where claims reach consumers or where national law extends it.
Is a JL-26 assessment a compliance certificate?
No. It reports a measured or estimated figure with its boundary and grade. It does not make anyone compliant with any law.
Sources
Keep reading
Take the next step.
Evidence and wording for AI energy claims as of November 2026: EU Directive 2024/825, the FTC Green Guides, boundaries, and measured vs modeled.