# Lead Green

> Lead Green is a public index of the GPU energy it takes to serve a million output tokens (watt-hours per million output tokens, Wh/M), by model, GPU and configuration, with grid carbon intensity to price it anywhere. Every figure comes from a named public source with a retrieval date. It is a Katonic project. Updated 27 September 2026.

Key readings (ML.ENERGY Leaderboard v3.0, Text Conversation workload, energy-optimal configuration at or under a 500 ms median inter-token latency, the leaderboard's default ceiling, which no configuration reaches):
- Lowest on the index: 13.0 Wh/M (Llama 3.1 8B Instruct, 1 × B200)
- Reference row: 63.5 Wh/M (Qwen 3 32B, 2 × H100)
- Highest on the index: 572.9 Wh/M (Llama 3.1 405B Instruct FP8, 8 × H100)
- The index spans 44x. The same model on the same GPUs moves 7.1x with concurrency (452 to 63 Wh/M for Qwen 3 32B on 2 x H100, running batch 8 to 256). Running batch = the average number of requests the server was decoding at once (vLLM num_requests_running); queued requests are not counted.
- Carbon: multiply Wh/M by the grid's gCO2e/kWh and divide by 1,000 for kg CO2e per billion tokens. Saudi Arabia 692, Serbia 668, Norway 30, World 474 gCO2e/kWh (Ember, 2024).
- Caveat: GPU energy only, measured at steady state on a busy server (the most energy-efficient configuration under a latency limit). Host, network, cooling and idle time are not counted. Every figure is a floor; a lightly used server sits well above it.
- Publisher and funder: Lead Green is run and paid for by Katonic, which sells a sovereign AI platform and benefits when buyers ask the questions on this site. It takes no money from hardware, cloud or model vendors.
- Contact: hello@leadgreen.ai, for corrections, measurements and data requests, answered within ten working days.
- Regulatory summaries on this site are as at 27 September 2026, for orientation, and are not legal advice. The EU AI Act asks providers of general-purpose AI models for training energy in MWh, held for the AI Office; no law asks anyone for serving energy per token.

## Pages
- [Home](https://leadgreen.ai/): what Lead Green is, why it exists, who it is for
- [Why it matters](https://leadgreen.ai/why/): nine reasons the number belongs in procurement, operations and reporting, with an RFP clause
- [The index](https://leadgreen.ai/index/): the full tables, method, sources ledger and change record
- [Models](https://leadgreen.ai/models/): one page per model, with every measured configuration, H100 against B200 and carbon by grid
- [H100 vs B200](https://leadgreen.ai/h100-vs-b200/): the 14 models measured on both cards, the gap per model and how much of it is the sweep
- [Grids](https://leadgreen.ai/grids/): one page per listed grid, the whole index priced on it, the trend since 2023 and the reporting line
- [Data API](https://leadgreen.ai/api/): read-only JSON, CORS open, CC BY 4.0
- [MCP server](https://leadgreen.ai/mcp/): tools for AI assistants (list_models, energy_per_million_tokens, estimate, compare_gpus, concurrency_curve, grid_intensity, gpu_hardware, sources)
- [Blog](https://leadgreen.ai/blog/): RSS at https://leadgreen.ai/feed.xml
- [Newsletter](https://leadgreen.ai/newsletter/)
- [Contribute](https://leadgreen.ai/contribute/): volunteer roles and the measurement protocol
- [About](https://leadgreen.ai/about/)

## Models (one page each; the figure is the lowest index row, GPU only, a floor)
- [Llama 3.1 8B Instruct](https://leadgreen.ai/models/llama-3-1-8b-instruct/): 13.0 Wh/M on B200
- [Qwen 3 8B](https://leadgreen.ai/models/qwen-3-8b/): 14.7 Wh/M on B200
- [Qwen 3 14B](https://leadgreen.ai/models/qwen-3-14b/): 15.2 Wh/M on B200
- [Qwen 3 30B A3B Instruct](https://leadgreen.ai/models/qwen-3-30b-a3b-instruct/): 15.7 Wh/M on B200
- [Gemma 3 12B](https://leadgreen.ai/models/gemma-3-12b/): 26.5 Wh/M on B200
- [Qwen 3 32B](https://leadgreen.ai/models/qwen-3-32b/): 39.1 Wh/M on B200
- [NVIDIA Nemotron Nano 12B V2](https://leadgreen.ai/models/nvidia-nemotron-nano-12b-v2/): 39.4 Wh/M on B200
- [NVIDIA Nemotron Nano 9B V2](https://leadgreen.ai/models/nvidia-nemotron-nano-9b-v2/): 40.9 Wh/M on B200
- [Gemma 3 27B](https://leadgreen.ai/models/gemma-3-27b/): 49.1 Wh/M on B200
- [Llama 3.1 70B Instruct](https://leadgreen.ai/models/llama-3-1-70b-instruct/): 52.4 Wh/M on B200
- [Llama 3.3 70B Instruct](https://leadgreen.ai/models/llama-3-3-70b-instruct/): 60.7 Wh/M on B200
- [Qwen 3 235B A22B Instruct](https://leadgreen.ai/models/qwen-3-235b-a22b-instruct/): 66.7 Wh/M on B200
- [Llama 4 Scout 17B 16E Instruct](https://leadgreen.ai/models/llama-4-scout-17b-16e-instruct/): 93.4 Wh/M on H100
- [Llama 4 Maverick 17B 128E Instruct FP8](https://leadgreen.ai/models/llama-4-maverick-17b-128e-instruct-fp8/): 134.5 Wh/M on H100
- [Qwen 3 235B A22B Instruct FP8](https://leadgreen.ai/models/qwen-3-235b-a22b-instruct-fp8/): 139.1 Wh/M on B200
- [DeepSeek V3.1](https://leadgreen.ai/models/deepseek-v3-1/): 168.4 Wh/M on B200
- [Llama 3.1 405B Instruct](https://leadgreen.ai/models/llama-3-1-405b-instruct/): 282.0 Wh/M on B200
- [Llama 3.1 405B Instruct FP8](https://leadgreen.ai/models/llama-3-1-405b-instruct-fp8/): 318.5 Wh/M on B200

## Data API (JSON)
- https://leadgreen.ai/api/v1/manifest.json: Sources, versions, retrieval dates, licences and this endpoint list.
- https://leadgreen.ai/api/v1/index.json: The index: 32 rows, one per model and GPU, energy-optimal under the latency limit, sorted ascending.
- https://leadgreen.ai/api/v1/models.json: Every model on the index with its slug, Hugging Face id, parameter counts and measured GPUs.
- https://leadgreen.ai/api/v1/models/{slug}.json: One model: every measured configuration (concurrency, parallelism, latency, power) and the best row per GPU.
- https://leadgreen.ai/api/v1/grid.json: Grid emissions intensity, gCO2e per kWh, 2024 annual average, 193 countries plus World.
- https://leadgreen.ai/api/v1/hardware.json: Manufacturer power ceilings, memory, MIG slice profiles and one embodied-carbon figure.
- https://leadgreen.ai/api/v1/per-query.json: AI Energy Score text-generation rows, December 2025 round: Wh per 1,000 queries at batch size 1 on H100 80GB GPUs (one card for classes A and B, several for C).
- https://leadgreen.ai/api/v1/openapi.json: OpenAPI 3.1 description of these endpoints.
- https://leadgreen.ai/api/v1/measurements.json: Contributed measurements accepted under the protocol, with contributor, affiliation, date, versions and MIG profile; never merged into index.json.
- https://leadgreen.ai/api/v1/schema/measurement.schema.json: JSON Schema for a contributed measurement file (protocol v0.2); the build validates every submission against it.
- https://leadgreen.ai/api/v1/changes.json: Every change to the published data or its description: date, summary, documents, reason and source version.
- https://leadgreen.ai/api/v1/index.csv: The index rows as CSV, for a spreadsheet.
- https://leadgreen.ai/api/v1/types.d.ts: TypeScript types for every document, generated from the documents.
- https://leadgreen.ai/api/v1/all.zip: Every JSON document of the current release in one zip (the same bytes as the dated release zip).
- https://leadgreen.ai/api/v1/releases/index.json: Every dated release of api/v1, each kept unchanged at releases/<date>/ with a zip of the whole release, for citation.

## Sources
- ML.ENERGY Leaderboard (The ML.ENERGY Initiative (started at SymbioticLab, University of Michigan)), 3.0 (released 2025-12-01); data compiled 2026-02-16, Apache-2.0 per the ml-energy/benchmark-v3 dataset card: https://ml.energy/leaderboard/
- Ember, Yearly electricity data, September 2026 release (file last modified 2026-09-22), CC-BY-4.0: https://ember-energy.org/data/yearly-electricity-data/
- NVIDIA Multi-Instance GPU User Guide, supported MIG profiles: https://docs.nvidia.com/datacenter/tesla/mig-user-guide/supported-mig-profiles.html
- NVIDIA product specifications: https://www.nvidia.com/en-us/data-center/a100/, https://www.nvidia.com/en-us/data-center/h100/, https://www.nvidia.com/en-us/data-center/h200/, https://www.nvidia.com/en-us/data-center/dgx-b200/
- NVIDIA, Product Carbon Footprint Summary for HGX B200: https://images.nvidia.com/aem-dam/Solutions/documents/HGX-B200-PCF-Summary.pdf
- AI Energy Score leaderboard (AI Energy Score (Hugging Face and Salesforce, with Neuralwatt, Cohere, Meta and Carnegie Mellon University)), December 2025 round: https://huggingface.co/spaces/AIEnergyScore/Leaderboard

## Licence
Compilation and site text: CC BY 4.0. Code: Apache 2.0. Third-party figures remain subject to their sources' terms, listed per source. Code: Apache 2.0.

## Markdown copies
Every page has a Markdown copy at its URL followed by index.md, generated from the same build and linked from the page head (rel=alternate, text/markdown).
- https://leadgreen.ai/about/index.md
- https://leadgreen.ai/api/index.md
- https://leadgreen.ai/blog/index.md
- https://leadgreen.ai/blog/what-a-million-tokens-costs-in-watt-hours/index.md
- https://leadgreen.ai/contribute/index.md
- https://leadgreen.ai/grids/australia/index.md
- https://leadgreen.ai/grids/bulgaria/index.md
- https://leadgreen.ai/grids/france/index.md
- https://leadgreen.ai/grids/germany/index.md
- https://leadgreen.ai/grids/hong-kong/index.md
- https://leadgreen.ai/grids/hungary/index.md
- https://leadgreen.ai/grids/index.md
- https://leadgreen.ai/grids/india/index.md
- https://leadgreen.ai/grids/ireland/index.md
- https://leadgreen.ai/grids/japan/index.md
- https://leadgreen.ai/grids/new-zealand/index.md
- https://leadgreen.ai/grids/norway/index.md
- https://leadgreen.ai/grids/qatar/index.md
- https://leadgreen.ai/grids/saudi-arabia/index.md
- https://leadgreen.ai/grids/serbia/index.md
- https://leadgreen.ai/grids/singapore/index.md
- https://leadgreen.ai/grids/slovakia/index.md
- https://leadgreen.ai/grids/sri-lanka/index.md
- https://leadgreen.ai/grids/taiwan/index.md
- https://leadgreen.ai/grids/the-philippines/index.md
- https://leadgreen.ai/grids/united-arab-emirates/index.md
- https://leadgreen.ai/grids/united-kingdom/index.md
- https://leadgreen.ai/grids/united-states/index.md
- https://leadgreen.ai/h100-vs-b200/index.md
- https://leadgreen.ai/index.md
- https://leadgreen.ai/index/index.md
- https://leadgreen.ai/mcp/index.md
- https://leadgreen.ai/models/deepseek-v3-1/index.md
- https://leadgreen.ai/models/gemma-3-12b/index.md
- https://leadgreen.ai/models/gemma-3-27b/index.md
- https://leadgreen.ai/models/index.md
- https://leadgreen.ai/models/llama-3-1-405b-instruct-fp8/index.md
- https://leadgreen.ai/models/llama-3-1-405b-instruct/index.md
- https://leadgreen.ai/models/llama-3-1-70b-instruct/index.md
- https://leadgreen.ai/models/llama-3-1-8b-instruct/index.md
- https://leadgreen.ai/models/llama-3-3-70b-instruct/index.md
- https://leadgreen.ai/models/llama-4-maverick-17b-128e-instruct-fp8/index.md
- https://leadgreen.ai/models/llama-4-scout-17b-16e-instruct/index.md
- https://leadgreen.ai/models/nvidia-nemotron-nano-12b-v2/index.md
- https://leadgreen.ai/models/nvidia-nemotron-nano-9b-v2/index.md
- https://leadgreen.ai/models/qwen-3-14b/index.md
- https://leadgreen.ai/models/qwen-3-235b-a22b-instruct-fp8/index.md
- https://leadgreen.ai/models/qwen-3-235b-a22b-instruct/index.md
- https://leadgreen.ai/models/qwen-3-30b-a3b-instruct/index.md
- https://leadgreen.ai/models/qwen-3-32b/index.md
- https://leadgreen.ai/models/qwen-3-8b/index.md
- https://leadgreen.ai/newsletter/index.md
- https://leadgreen.ai/why/index.md