A Katonic project, edited by Prem Naraindas
Lead Green: the energy behind the token
Lead Green is a public index of the GPU energy it takes to serve a million AI tokens: watt-hours per million output tokens, by model, GPU and how busy the server is, with the grid carbon to price it anywhere. Every figure comes from a named public measurement or a manufacturer's specification, dated and linked. The data is open, there is a JSON API and an MCP server, and anyone with a GPU will be able to add a measurement.
Open the index at leadgreen.ai The dataset
Why a sovereignty person runs an energy index
Every on-premise deployment Katonic has done meets the same question on day one: a fixed power feed, a fixed GPU count, and a customer who wants to know how many tokens that buys. The hyperscaler never has to answer it, because the meter is somewhere else. A government or a bank running its own AI has to. The index started as the answer we wished existed.
It is also the honest half of the sovereignty argument. The Sovereignty Test asks who holds the keys and who can switch the system off. Lead Green asks what it costs, in energy, to keep it on. Control without a power budget is a slogan.
What is in it
- One unit. Watt-hours per million output tokens, written Wh/M. GPU only, on a busy server. Every row is a floor: the number a buyer should ask for is the one measured on their own machine.
- The index. At the 19 October 2026 release: 32 rows across 18 open-weight models and two GPU generations, each with the concurrency it was measured at and a link to the source.
- Grid carbon. Carbon intensity for 194 countries and regions, so a Wh/M figure becomes grams of carbon in Riyadh, Sydney or Oslo.
- For machines. A versioned JSON API, an MCP server so an agent can look a figure up, a Python package, and the full dataset on Hugging Face, Kaggle and Zenodo.
- A change record. Every dated release of the API is kept unchanged. Corrections go through an editor, with the method published.
Three numbers from the index
| Row | Wh/M | Configuration |
|---|---|---|
| Lowest floor | 13.0 | Llama 3.1 8B Instruct on one B200, running batch 2,043 |
| Reference floor | 63.5 | Qwen 3 32B on two H100s, running batch 256. The carbon table is priced on this row. |
| Highest floor | 572.9 | Llama 3.1 405B Instruct FP8 on eight H100s, running batch 255 |
The index spans 44 times from its lowest row to its highest, and the same model on the same two GPUs moves 7.1 times depending on how busy the server is. A million output tokens is roughly three quarters of a million words of English. Figures as of the 19 October 2026 release; the live index is the record.
Where it lives
- The site and the index: leadgreen.ai
- The dataset: Hugging Face, leadgreen/green-token-index
- The archived release with a DOI: 10.5281/zenodo.23478909 on Zenodo
- The entity: Wikidata Q141778099
- The code and site source: GitHub, katonic-dev/leadgreen-ai
- The updates: Lead Green on LinkedIn
Who runs it
Lead Green is a Katonic project, not a separate company and not a product. Katonic AI publishes it; I edit it. Corrections, contributions and data requests go to hello@leadgreen.ai and are answered within ten working days. The data is licensed CC BY 4.0: use it, cite it, improve it.
How to cite
Lead Green, a Katonic project. Green Tokens index, release 2026-10-19. https://leadgreen.ai/index/, retrieved [date]. CC BY 4.0. DOI 10.5281/zenodo.23478909.
What I talk about when I talk about it
Tokens per watt-hour is becoming a procurement number, the way price per token already is. For a nation or an enterprise with a fixed power feed, it is the number that decides how much AI the building can run. That makes it a sovereignty metric as much as an efficiency one, and it belongs in the same conversation as the keys and the off switch. It is a standing topic in my talks and in the newsletter.