Grids · Oceania · a market Katonic sells in or expects to
New Zealand: the carbon of AI inference per billion tokens
New Zealand's grid ran at 112 gCO2e per kWh in 2024 on Ember's yearly figures, 4.2× cleaner than the world average of 474, and 26th of the 193 countries Ember covers, counting from the cleanest. A billion output tokens of the index's reference row (Qwen 3 32B, 2 × H100, 63.5 Wh/M) served on it carries about 7.1 kg CO2e of GPU-side emissions; the index's lowest row carries 1.5 kg and its highest 64.1 kg. Energy per token does not change with geography; carbon per token does.
Area New Zealand (NZL, NZ). Basis lifecycle, location-based, annual average. Data grid.json. Updated 27 September 2026.
Carbon per token is energy per token times the grid. The energy figures here are GPU-only floors from a busy server; a lightly loaded deployment sits above them, and so does anything that includes the host, the network and the cooling. Wh/M × gCO2e per kWh ÷ 1,000 gives kg CO2e per billion output tokens.
Trend
New Zealand's grid, year by year
From 2023 to 2025 the intensity rose 6 percent, from 87.4 to 92.7 gCO2e per kWh. The table uses 2024, the latest year with a value for every listed grid; on the 2025 figure the reference row would carry 5.9 kg per billion tokens.
| Year | g CO2e per kWh | kg CO2e per billion tokens, reference row |
|---|---|---|
| 2023 | 87.4 | 5.5 |
| 2024 | 111.9 | 7.1 |
| 2025 | 92.7 | 5.9 |
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. A GPU server draws much the same power around the clock, while an annual average weights each hour by its generation, so on a grid whose clean supply peaks at midday a flat load sees a higher figure than the average.
Every model
The index priced on New Zealand's grid
Each model's lowest index row, GPU only, at the most energy-efficient configuration measured, multiplied by 112 gCO2e per kWh. The spread from the lowest row to the highest is 44×, and it is entirely the model, the GPU and the load; the grid multiplies all of them alike.
| Model | GPUs | Wh / M | kg CO2e per billion output tokens |
|---|---|---|---|
| Llama 3.1 8B Instruct | 1 × B200 | 13.0 | 1.5 |
| Qwen 3 8B | 1 × B200 | 14.7 | 1.6 |
| Qwen 3 14B | 1 × B200 | 15.2 | 1.7 |
| Qwen 3 30B A3B Instruct | 1 × B200 | 15.7 | 1.8 |
| Gemma 3 12B | 1 × B200 | 26.5 | 3.0 |
| Qwen 3 32B | 1 × B200 | 39.1 | 4.4 |
| NVIDIA Nemotron Nano 12B V2 | 1 × B200 | 39.4 | 4.4 |
| NVIDIA Nemotron Nano 9B V2 | 1 × B200 | 40.9 | 4.6 |
| Gemma 3 27B | 1 × B200 | 49.1 | 5.5 |
| Llama 3.1 70B Instruct | 2 × B200 | 52.4 | 5.9 |
| Llama 3.3 70B Instruct | 2 × B200 | 60.7 | 6.8 |
| Qwen 3 235B A22B Instruct | 4 × B200 | 66.7 | 7.5 |
| Llama 4 Scout 17B 16E Instruct | 4 × H100 | 93.4 | 10.4 |
| Llama 4 Maverick 17B 128E Instruct FP8 | 8 × H100 | 134.5 | 15.0 |
| Qwen 3 235B A22B Instruct FP8 | 2 × B200 | 139.1 | 15.6 |
| DeepSeek V3.1 | 8 × B200 | 168.4 | 18.8 |
| Llama 3.1 405B Instruct | 8 × B200 | 282.0 | 31.5 |
| Llama 3.1 405B Instruct FP8 | 4 × B200 | 318.5 | 35.6 |
Every configuration of every model, not only the lowest row, is on the model pages; the MCP server prices any of them on New Zealand's grid for a token count you give it.
Reporting
Which line it goes on
For GPUs you run in New Zealand, the electricity is your own consumption: it goes into your ESRS E1-5 energy total and its emissions into E1-6 as Scope 2, location-based, which is the method this page's figure matches. The market-based figure nets your contracted supply and needs your own contracts. Tokens bought from a hosted API are the provider's Scope 2 and your Scope 3 category 1, and sit outside your E1-5 line; the why page lists what to ask the provider. Ember's figure is a lifecycle intensity for comparing places, not a Scope 2 combustion factor to file as reported.
Questions
Asked about AI carbon in New Zealand
What is the grid carbon intensity of New Zealand?
112 gCO2e per kWh in 2024, on Ember's yearly electricity data: a lifecycle, location-based figure for the country's own generation, 4.2× cleaner than the world average of 474. Ember's 2025 figure is 92.7.
How much CO2 does a million AI tokens produce in New Zealand?
At the index's reference row (Qwen 3 32B on 2 × H100, 63.5 Wh per million output tokens, GPU only), about 7.1 grams of CO2e per million tokens, or 7.1 kg per billion, on New Zealand's 2024 grid. Across the index the range is 1.5 to 64.1 kg per billion tokens. A lightly loaded server sits above these floors, and host, network and cooling are not counted.
Is running AI in New Zealand cleaner than the world average?
New Zealand's 2024 intensity of 112 gCO2e per kWh is 4.2× cleaner than the world average of 474 (474 gCO2e per kWh). The same tokens on the same GPUs produce the same energy figure anywhere; only the carbon figure moves with the grid.
Can I use this figure for my Scope 2 report in New Zealand?
As a comparison, not as a filing. Ember's intensity is a lifecycle figure on net generation, not a Scope 2 combustion factor, and it is location-based; a market-based figure needs your own contracted supply. Tokens bought from a hosted API land in your Scope 3 category 1, not your Scope 2. The why page explains both.
Where do these figures come from?
Grid intensity: Ember, Yearly electricity data, September 2026 release (2024; 2023, 2024, 2025 shown), CC BY 4.0, retrieved 26 September 2026. Energy per token: the ML.ENERGY Leaderboard v3.0 (data compiled 16 February 2026). Carbon is Wh/M × gCO2e per kWh ÷ 1,000.
Sources
Where every figure came from
Grid intensity: Ember, Yearly electricity data, September 2026 release, emissions intensity of total generation, 2023, 2024, 2025 (CC BY 4.0), retrieved 26 September 2026. Energy per token: ML.ENERGY Leaderboard version 3.0 (released 1 December 2025; data compiled 16 February 2026; retrieved 26 September 2026), The ML.ENERGY Initiative, started at SymbioticLab, University of Michigan. Text Conversation task, 328 stable configurations. Energy per token is steady-state GPU energy divided by the output tokens generated in the window, measured with Zeus under vLLM 0.11.1 on H100 80GB HBM3 and B200 servers; energy per request is that figure times the average output length. Each index row is the lowest-energy configuration for its model and GPU at or under a 500 ms median inter-token latency, the leaderboard's own default ceiling, which no configuration reaches. Apache 2.0 per the ml-energy/benchmark-v3 dataset card. Cite Chung et al., The ML.ENERGY Benchmark, NeurIPS 2025 Datasets and Benchmarks.
Cite as. Lead Green, a Katonic project. Green Tokens: GPU energy per million output tokens. https://leadgreen.ai/index/, release 2026-09-27, retrieved [date]. © 2026 Katonic AI Technology Ltd. Licensed CC BY 4.0, offered as-is and without warranty; state whether you changed any rows or columns. Underlying data: ML.ENERGY Leaderboard v3.0 (The ML.ENERGY Initiative; Apache 2.0 per the ml-energy/benchmark-v3 dataset card; cite Chung et al., The ML.ENERGY Benchmark, NeurIPS 2025 Datasets and Benchmarks, arXiv 2505.06371); Ember, Yearly electricity data, September 2026 release (CC BY 4.0); AI Energy Score (Apache-2.0 as tagged on its Space; results reproduced as facts with attribution); NVIDIA figures reproduced as facts from the documents named in the sources ledger.
This page is generated from the same dated snapshots as the index; no figure on it is typed by hand. All 193 areas are in grid.json; the 22 with a page are the markets Katonic sells in or expects to, plus seven reference grids.