LLM Carbon Calculator

As AI usage grows across organisations, so does the question of its environmental cost. This calculator estimates the CO₂e emissions associated with running large language models — based on how many tokens you process each month, the carbon intensity of the electricity grid powering the data centre, and optional factors like hardware efficiency and embodied chip emissions.

How to use it: Set your monthly token volume in Step 1, then choose a grid profile matching your provider's data centre location in Step 2. The race track below updates in real time, showing how your footprint compares across three broad model tiers. Click any car to see a full breakdown. Advanced parameters (PUE, training amortisation, embodied carbon) are available if you want to dig deeper.

1 Monthly Usage Level

Small team

~1.5M tokens / month

3–5 solicitors actively using AI across matters

Monthly token volume 300k tokens

10k100k1M10M100M500M

2 Grid Carbon Intensity

Carbon intensity of electricity 233 gCO₂e/kWh

CO₂ equivalent per kWh consumed at the data centre. Varies widely by country and whether the provider has purchased renewable energy certificates.

🇬🇧 Equivalent to United Kingdom average grid (233 gCO₂e/kWh)

3 Advanced Options

PUE, input/output ratio, embodied carbon ▾

Data-centre PUE (includes cooling) 1.20×

Power Usage Effectiveness — the ratio of total facility energy to IT equipment energy. This is the mechanism by which cooling is already included in this model: a PUE of 1.2 means cooling and power distribution add 20% on top of raw compute energy. Hyperscaler leaders (Google, Microsoft) run ~1.09–1.10; mid-tier cloud ~1.2–1.4; legacy data centres ~1.5–2.0. Water-based evaporative cooling tends to achieve lower PUE at the cost of water consumption, which is not captured here.

Input : output token ratio 3 : 1

Input (prompt) tokens consume ~30% the energy of output tokens per token, because generation is memory-bandwidth-bound while prefill can be batched efficiently. Higher ratios reduce effective energy per total token.

Model training carbon (amortised) 5%

Training a frontier model emits thousands of tonnes of CO₂e — but that cost is shared across every inference call made during the model's lifetime. At population scale, inference now accounts for ~80–90% of a model's total lifecycle energy, making training a relatively small per-token overhead. 5% is a reasonable mid estimate; set to 0% if you believe you are using a model trained primarily on renewables.

GPU/chip embodied carbon (amortised) 15%

Manufacturing GPUs (chip fabrication, packaging, logistics, end-of-life) emits carbon independently of electricity use. Lifecycle analyses suggest embodied hardware emissions add roughly 24–35% to the total footprint of an LLM deployment when amortised over hardware lifetime (~3–6 years). Data centre building construction is a much smaller additional term (~2–5%) and is not modelled separately. Set to 0% to see operational-only emissions.

// Monthly carbon footprint — equivalent km driven per model tier

Click a car to see the full breakdown for that model tier.

Tiny / Small e.g. Llama 3.1 8B, Haiku

🚗

—

Mid / Standard e.g. GPT-4o, Sonnet

🚗

—

Large Frontier e.g. GPT-4.5, o3, Opus

🚗

—

Distances are log-scaled · 170 gCO₂e/km (UK avg petrol car) · click any car for details

What this calculator covers: The marginal operational cost of your inference usage — the electricity consumed by GPUs and servers to process your tokens (via J/token), scaled by data centre overhead (PUE, which already includes cooling), multiplied by the grid's carbon intensity. On top of this, two amortised embodied terms are optionally added: a share of the model's training carbon, and a share of GPU/chip manufacturing emissions spread over hardware lifetime.

What is not included: Water consumption (a real impact but measured in litres, not CO₂e, with no agreed per-token conversion); data centre building construction (small, ~2–5%, amortised over ~20 years); network transmission energy (typically <5% of inference energy); and the embodied carbon of your own device. These omissions mean the figure shown is a lower bound on the true lifecycle footprint. The largest single excluded category at the individual-user level is likely chip embodied carbon — use the slider above to include an estimate.

↗ Data Sources

8 peer-reviewed papers & datasets underpinning the model ▾

inference energy

From Words to Watts

Samsi et al., IEEE HPEC 2023

GPU benchmark measuring energy per token across model sizes. Llama 65B measured at ~4 J/token on A100 hardware under realistic batch conditions.

arxiv.org/abs/2310.03003

frontier model estimate

How Much Energy Does ChatGPT Use?

Josh You, Epoch AI, February 2025

Bottom-up H100 TDP model estimating ~0.3 Wh per GPT-4o query (~500 output tokens), yielding ~2.2 J/token. Independently corroborated by OpenAI's own disclosure.

epoch.ai/gradient-updates/…

inference energy · benchmarks

TokenPowerBench

Niu et al., AAAI 2026

Systematic benchmark across Llama (1B–405B), Falcon, Qwen, Mistral. Finds J/token scales ~7.3× from 1B→70B parameters — sublinear vs parameter count due to memory-bandwidth constraints.

arxiv.org/abs/2512.03024

frontier model energy

How Hungry is AI?

Jehham et al., arXiv May 2025

Per-query energy estimates for GPT-4o (0.43 Wh short query), o3, and GPT-4.5 across prompt lengths. Confirms inference now accounts for up to 90% of model lifecycle energy.

arxiv.org/abs/2505.09598

embodied carbon

LLMCarbon: End-to-end Carbon Footprint of LLMs

Ahmad et al., ICLR 2024

Models both operational and embodied (chip fabrication, packaging, logistics) carbon. Finds embodied emissions add ~24–35% to total LLM footprint when amortised over hardware lifetime. Validates against Google's published LLM carbon figures.

openreview.net/pdf?id=aIok3ZD9to

embodied carbon · hardware lifecycle

Life-Cycle Emissions of AI Hardware

Google, arXiv 2025

Lifecycle analysis of Google TPU generations including chip manufacturing, data centre construction, and transportation. Shows manufacturing emissions per chip increase with newer fabrication nodes despite efficiency gains — used to calibrate the chip embodied carbon range.

arxiv.org/abs/2502.01671

2024 Global Data Center Survey

Uptime Institute, 2024

Global average PUE of 1.56; hyperscaler leaders achieving 1.08–1.09. Used to calibrate the PUE slider default of 1.2 (reflecting mid-tier cloud providers).

grid carbon intensity

Electricity Maps / Ember Climate

Ember Climate & electricitymaps.com

Country-level carbon intensity figures (gCO₂e/kWh) for grid presets. UK default of 233 gCO₂e/kWh reflects the 2023–24 annual average. Values shift year to year with renewable build-out.

ember-climate.org/data/…

LLM Carbon Footprint Calculator