Every number behind the badge, published.
Offset Verified is a substantiated claim, not a slogan. This page is the methodology: how token usage becomes energy, water, CO2e, and a dollar restitution target, with the coefficients in the open.
What Offset Verified means.
The organization measured its AI usage from provider records, Eniyan converted that usage into a resource restitution target using the published method on this page, and the organization funded evidence-backed offsets, reviewed by our team, that fully cover the cumulative target.
Carbon neutral, net zero, or environmentally harmless. The badge verifies a specific, bounded act of restitution: covering the estimated replacement cost of the power and water the organization's AI workloads consumed. The dollar target is a utility replacement cost, not a carbon-credit price.
From tokens to a dollar target, in five steps.
Eniyan never observes your inference. The inputs are the usage exports your AI providers already give you.
- 1Usage: your team uploads provider usage exports (token counts by model and month). Each model is mapped to a size class.
- 2Energy: tokens convert to watt hours using the per-class rates below, output tokens weighted 4:1 over input, then multiplied by the 1.2 datacenter overhead factor.
- 3Water: facility energy converts to liters at 1.0 liters per kWh.
- 4Emissions: energy converts to CO2e at the period's grid intensity (370 gCO2e per kWh when no regional rate is on file).
- 5Restitution target: the energy and water are priced at current utility rates averaged across the eight highest-concentration US datacenter states, giving a dollar figure that would replace what was consumed.
The coefficients, verbatim.
Conservative point estimates, versioned in code. Every footprint row is stamped with the version that computed it, and version bumps never rewrite finalized periods.
| Model class | Wh per 1,000 tokens | Typical members |
|---|---|---|
| Frontier | 2.5 | the largest reasoning models |
| Large | 1.0 | mainstream flagship models |
| Medium | 0.35 | mini, flash, and haiku class models |
| Small | 0.10 | nano and embedding models |
Output tokens are weighted four times input tokens around each class blend (reference mix 20% input, 80% output).
Power usage effectiveness multiplier applied to model energy.
On-site water use effectiveness per facility kWh.
US average, consumption based. Used when no period rate rows exist.
Commercial retail electricity prices pulled monthly from the US EIA open data API; water from curated municipal rate schedules.
Averaged across the US states with the highest datacenter concentration.
The rules that keep the badge honest.
Every verified organization gets a public badge page showing its status, the import-backed totals behind it, and this methodology. If a badge lapses, the public page says so. Talk to us about enabling Sustainability for your organization.