Lead Green opens a GPU energy index that AI assistants can query

Lead Green has launched a public index for a question token bills usually leave unanswered: how much GPU energy does a million output tokens take? The Katonic-funded project combines a browsable site, downloadable data, a read-only JSON API and MCP tools that let an assistant calculate estimates from named measurement sources.

The October 19 launch is a new publication and access point, not a new set of GPU experiments. The change log dates the project’s first build to September 26 and explicitly says the launch changes neither the data nor its description. Its Hugging Face snapshot appeared on October 19 at 15:41:20 UTC.

One unit, with a narrow boundary

The reviewed snapshot contains 32 index rows covering 18 models. Each row selects the lowest-energy measured configuration for a model and GPU under a median inter-token latency ceiling of 500 milliseconds. The underlying inference measurements come from ML.ENERGY Leaderboard v3.0 data compiled on February 16, 2026; the project preserves that date instead of presenting it as October testing.

The index’s rounded range is 13.0 watt-hours per million output tokens for Llama 3.1 8B Instruct on one B200, through 572.9 for Llama 3.1 405B Instruct FP8 on eight H100s. Those are different models, hardware and loads, not an equal-quality contest or proof that one model is universally more efficient.

The boundary matters more than the ranking. The figures cover steady-state GPU energy on a busy server, excluding host systems, networking, cooling and idle time. Lead Green calls them floors: a lightly used deployment can consume more per token. Annual grid-emissions data from Ember supplies a separate carbon conversion; it does not change the GPU energy measurement or establish the real-time emissions of a particular request.

An assistant interface to published data

The MCP component is a local stdio server, leadgreen-mcp 0.2.0, rather than a hosted HTTP endpoint. The documentation requires Python 3.10 or later and offers installation from the repository’s leadgreen-mcp subdirectory, or a uvx command. It documents Claude Desktop, Claude Code and Cursor configurations; other clients need compatible MCP support.

Its eight tools include model listing, energy estimates, GPU comparisons, concurrency curves, grid intensity and source lookup. The public JSON API needs no key. According to the project, the server fetches published JSON with HTTPS GET requests and performs calculations locally, without transmitting tool questions, token counts or country inputs. It can fall back to a bundled snapshot or read a local data folder; responses identify their data source and update date. An assistant’s own model-provider usage is separate from this data access.

The compilation and site text are CC BY 4.0, while the code is Apache-2.0 and third-party source terms remain applicable. Katonic discloses that it funds the project and sells a sovereign AI platform. For engineers and procurement teams, the useful addition is a common, inspectable baseline and an assistant-friendly way to query it. It is not a measurement of every hosted AI API or a substitute for testing the actual load and full-system power of a deployment.