Green Tokens

Two generations

The same model on H100 and B200

14 models were measured on both GPUs. The newer card's advantage runs from 4.55× on Qwen 3 235B A22B Instruct to 1.05× on NVIDIA Nemotron Nano 9B V2. The gap also depends on which configurations were swept. The 4.55× pair compares an H100 row that was run only at eight cards and a cap of at most 256 requests with a B200 row at a running batch of 3,765 on four cards; at the same cap of 256 the B200 row reads 197.5 Wh/M against the H100's 303.5, a gap of 1.54×. The two Nemotron pairs compare runs whose average answers are 3.2 and 2.4 times longer on the B200 (462 and 510 tokens on the H100 against 1,464 and 1,242).

At a tighter ceiling. At a 50 ms ceiling the comparison shifts: 14 models still have a row on both cards, and the H100 uses less energy per token than the B200 on two of them (Llama 3.1 8B Instruct, 19.5 Wh/M on 1 × H100 against 19.6 on 1 × B200; Qwen 3 235B A22B Instruct FP8, 306.7 Wh/M on 8 × H100 against 408.9 on 2 × B200).

A card that draws more power can still use less energy per token, because it finishes more tokens per second. Power tells you the size of the feed. Energy per token tells you what the feed produced. The pairs, with card counts and running batches, are on the H100 vs B200 page.

Concurrency

The idle server is the expensive one

Energy per token depends on how many requests the server is handling at once. Below is the same Qwen 3 32B on the same two H100 cards, all nine measured configurations, from a running batch of 8 to the index row at 256. Energy per token falls 7.1×, from 452 to 63.5 Wh/M, while median inter-token latency stays at 44 ms or under. Power rises 39 percent, from 850 W to 1,178 W, while throughput rises 9.9×, from 522 to 5,155 tokens per second. Energy per token falls because the second effect dwarfs the first. Past the best row the curve turns: at a cap of 384 the same pair of cards reads 64.6 Wh/M at 55 ms.

A departmental deployment that serves a handful of people at a time lives at the top of these tables, not the bottom. In the editor's view, consolidating traffic onto fewer, busier cards is a lever every operator already holds, it costs nothing in hardware, and on these tables it moves energy per token further than a change of GPU generation does.

The Llama 3.1 8B ladder has eleven measured configurations. Its power rises 41 percent, from 486 W to 688 W, while energy per token falls 6.5×, from 126 to 19.5 Wh/M; past its best row, at a cap of 768, it reads 19.8 Wh/M at 63 ms.

Cap is the configured maximum batch (vLLM's max_num_seqs). Running batch (avg) is the average number of requests the server was decoding at once over the steady-state window (vLLM's num_requests_running); queued requests are not counted, and the benchmark keeps a queue at all times. The highlighted row is the index row.

Same source and workload as the index. Each row is one benchmarked configuration; the index takes the best of them. Every model's full curve is in its API document.

Two tasks

Long answers also shrink the batch

The same leaderboard also runs a Problem Solving task, whose answers are far longer. For the 11 model and GPU pairs measured on both tasks, Problem Solving costs 3 to 41 times the energy per response of Text Conversation, because its answers run 5,361 to 8,969 tokens against 462 to 1,464, and because the long outputs cut the batch the server can hold, so energy per token itself is 2.2 to 3.3 times the Text Conversation figure for the Qwen and DeepSeek pairs, and 0.7 to 1.0 times it for the four Nemotron pairs.

Both ratio columns are Problem Solving over Text Conversation for the same model and GPU. Energy per token is the fairer comparison across workloads; energy per response is what a bill sees. Same source as the index.

Carbon

Same tokens, different grid

Energy per token does not change with geography. Carbon per token does. This table prices the reference row (Qwen 3 32B, 2 × H100, 63.5 Wh/M) against each grid's average emissions intensity for 2024. A billion output tokens on the Saudi grid carries about 43.9 kg CO2e of GPU-side emissions. The same billion tokens on the Norwegian grid carries 1.9 kg.

Which grids. The 22 grids are 15 markets Katonic sells in or expects to, plus seven reference grids. Ember's 2024 figures run from Ethiopia at 23.5 to Turkmenistan at 1,306.2 gCO2e per kWh, a spread of 55.6×; all 193 areas plus World are in grid.json, and each listed grid has its own page with the whole index priced on it. Area names are shortened from Ember's labels.

What the figure is. Grid figures are Ember's yearly emissions intensity (CC BY 4.0), September 2026 release. Lifecycle emissions intensity: CO2e per kWh of net generation, built from per-technology lifecycle factors (IPCC AR5 WG3 Annex III midpoints, all gases at GWP-100), so upstream methane, supply chains and plant manufacture are included. Production-based: imports and grid losses are excluded. A location-based figure for comparing places, not a Scope 2 combustion factor to file as reported. The table is location-based in the ESRS E1-6 sense, so what it prices belongs in the Scope 2 location-based line of E1-6, not in the E1-5 energy line. The other method, market-based, nets the buyer's contracted supply (certificates and power purchase agreements) and needs the buyer's own contracts.

Why 2024. It is the latest year with a value for every grid listed. 93 of Ember's 193 areas already have a 2025 value, but Saudi Arabia, United Arab Emirates and Hong Kong do not yet; grid.json carries 2023 to 2025 per area where present, with the latest year named. A flat load. A GPU server draws much the same power around the clock, while an annual average weights each hour by how much was generated in it, so on a grid whose clean supply peaks at midday a flat load sees a higher figure than the average, and on one with steady clean supply much the same. Hourly intensity, contracted supply, on-site generation and facility overhead are not counted here. Multiply any row of the index by the grid figure and divide by 1,000 to price a different model.

Hardware

What a card can draw, and how it splits

Manufacturer ceilings, for scale. A card at the index's best rows runs close to these numbers. The concurrency tables above show what that draw earns.

The A100 80GB uses the same six shapes as the H100 column (the seventh, a 1g media-extension variant, is omitted for all four cards), and the B200 180GB has the same shapes at its own sizes (all four cards are in the API). NVIDIA does not publish power draw per MIG slice, and neither public benchmark on this page measures a sliced card. That gap is the first measurement Lead Green will add.

Sources: NVIDIA product specifications for A100, H100 and H200; NVIDIA Product Carbon Footprint Summary for HGX B200 (July 2025) for the B200 ceiling; NVIDIA Multi-Instance GPU User Guide, supported profiles, last updated 11 September 2026. The "+me" media-extension variants of the 1g profiles are omitted.

Per query

Batch size one, a different question

The AI Energy Score measures something else: GPU watt-hours for 1,000 text-generation queries, one query at a time, on NVIDIA H100 80GB GPUs: one card for class A and B models, several cards for class C. A query is one prompt from the benchmark's 1,000-prompt text-generation set, run at batch size 1 with the response fixed at ten generated tokens for non-reasoning models, averaged over ten passes. It is a controlled probe, not a serving workload, so its watt-hours are not comparable with the index and are not converted. It is shown because the ranking within a class is useful.

Only the December 2025 round is shown, all 22 of its rows, and they are in the API. Lead Green does not mix it with the February 2025 round because the benchmark harness changed between them (the February round ran on Optimum Benchmark, the December round on Neuralwatt's AI Energy Benchmarks); the project itself treats the watt-hour values as stable across rounds and recalibrates only the stars. The 70B figures people quote from the two rounds are two different models: Meta-Llama-3-70B at 1,719.7 Wh in February 2025 and DeepSeek-R1-Distill-Llama-70B at 49.5 Wh in December, while Apertus-70B reads 697.5 Wh inside the December round.

Stars are the AI Energy Score's rating within a class for this round: the class's energy range is cut into five equal bands and the lowest band is five stars. They do not compare across classes: zai-org/GLM-4.5-Air, five stars in class C at 220.5 Wh, uses 4.0 times the energy of LGAI-EXAONE/EXAONE-4.0-32B, a one-star class B model at 54.5 Wh. Class A fits one consumer GPU, B one H100, C more than one H100. Source: AI Energy Score (Hugging Face and Salesforce, with Neuralwatt, Cohere, Meta and Carnegie Mellon University), retrieved 26 September 2026.

Method

What is counted, and what is not

Own measurements

None yet

Contributed measurements

Rows measured by volunteers

None yet.

Rows measured under the protocol and accepted by two reviewers appear here with the contributor's name, and in measurements.json. They are never merged into the index table above.

Cite as

How to credit this index

Sources

Every figure, and where it came from

1. ML.ENERGY Leaderboard, The ML.ENERGY Initiative (started at SymbioticLab, University of Michigan)

   Index, two-generation table, concurrency tables, the workload comparison and the measured GPU power figures. Text Conversation, all 328 stable configurations; Problem Solving, all 187. The leaderboard publishes the runs that passed its stability filter as JSON in its repository; every run, including the unstable ones, is in the gated Hugging Face dataset ml-energy/benchmark-v3. The leaderboard repository the JSON comes from carries no licence file; the benchmark data is published under Apache 2.0 on the dataset card and in the NeurIPS 2025 paper, and the JSON is reproduced here as published measurements on that basis while the ML.ENERGY team is asked to confirm that the licence covers it. Cite: Chung et al., The ML.ENERGY Benchmark: Toward Automated Inference Energy Measurement and Optimization, NeurIPS 2025 Datasets and Benchmarks, arXiv 2505.06371.

   Version 3.0, released 1 December 2025; data compiled 16 February 2026 (the leaderboard's index.json). Retrieved 26 September 2026.
2. Ember, Yearly electricity data

   Carbon table and the grid API. Emissions intensity of total generation, gCO2e per kWh, 2024. CC BY 4.0.

   September 2026 release. Retrieved 27 September 2026.
3. NVIDIA Multi-Instance GPU User Guide, Supported MIG Profiles

   MIG profile table.

   Last updated 11 September 2026. Retrieved 27 September 2026.
4. NVIDIA product specifications: A100, H100, H200, DGX B200

   Maximum TDP and memory in the hardware table; the DGX B200 system power figure used on the "why" page.

   Retrieved 27 September 2026.
5. NVIDIA, Product Carbon Footprint Summary for NVIDIA HGX B200

   B200 maximum TDP (configurable up to 1,000 W per GPU) and the 2,274 kg CO2e cradle-to-gate figure in the method.

   July 2025. Retrieved 27 September 2026.
6. AI Energy Score leaderboard (Hugging Face and Salesforce, with Neuralwatt, Cohere, Meta and Carnegie Mellon University), with its documentation

   Per-query table. Text generation task, December 2025 round, all 22 rows. Apache-2.0 as tagged on the Space; the project states no separate licence for the results file, which is reproduced as facts with attribution.

   Retrieved 27 September 2026.
7. OpenAI, What are tokens and how to count them

   The rough words-per-token conversion in the opening paragraph.

   Retrieved 27 September 2026.

Snapshots of every data file used, as retrieved, are kept alongside the page so each figure can be recomputed. Corrections are welcome and are recorded below.

Changes

Record

• 27 Sep 2026Corrections after a nine-reader review, before launch. The ML.ENERGY data is dated to its compile date, not a banner update; the latency ceiling is the leaderboard's own 500 ms (no row moved); "requests in flight" is renamed running batch; full concurrency ladders; all AI Energy Score December rows; a two-task table; the carbon basis, grid years and display names; seven-decimal joules; additive API fields (data_version, row_id, repeat, p95_itl_ms, avg_running_batch, boundary, years, latest_year, iso2, display_name, source_label); changes.json, measurements.json, the measurement schema, a dated release archive, CSV, zip and TypeScript types.
• 26 Sep 2026First build. 32 index rows from ML.ENERGY v3.0 (data compiled 16 February 2026), grid intensities from Ember (2024), NVIDIA specifications, AI Energy Score December 2025 round. The API, the MCP server and the first blog post.