TRANSPARENCY & SOURCES

How We Calculate AI's Environmental Footprint

Our estimates are grounded in official provider disclosures, peer-reviewed academic research, and a transparent estimation framework. Every figure shows its confidence level.

3
Source tiers
4
Verified / peer-reviewed
Nov 2026
Last data review

Our Approach

Know Your Compute provides transparent, data-backed estimates of the environmental cost of AI inference.

We prioritise official disclosures from providers, then peer-reviewed academic research, then transparent estimation. We never present a guess as a fact — every figure on this site shows its confidence level.

All calculations cover inference only (the energy used when you interact with a model), not training. We focus on per-query impact because inference now accounts for 80–90% of AI's total energy consumption.

Data Source Tiers

Verified

Tier 1 — Verified

Official disclosures from providers:

Estimated

Tier 3 — Estimated

Our estimation framework, based on:

  • CLEER (Sustainable AI Group, September 2026) — per-model estimates for 51 current models, benchmarked on open weights and projected to closed ones
  • GPU system power specifications (NVIDIA H100/B200/B300, Google TPU)
  • API performance benchmarks (latency, throughput)
  • Regional environmental multipliers (PUE, WUE, CIF)

The Estimation Formula

For models without official disclosures, we estimate energy per query as:

E_query = (P_gpu × N_gpu × T_inference × PUE) / 3600
E_query
Energy per query in watt-hours (Wh)
P_gpu
GPU system power in watts, including server overhead (~1.72× the GPU's rated TDP)
N_gpu
Number of GPUs needed per query (derived from model size and memory requirements)
T_inference
Inference time in seconds (from API benchmarks or throughput estimates)
PUE
Power Usage Effectiveness of the data centre (1.0 = perfect, typical is 1.1–1.35)

CO₂ and water are then derived from energy using regional environmental data:

CO₂ (grams) = E_query × CIF / 1000
Water (mL) = E_query × (site_WUE + source_WUE)

Regional Environmental Data

Environmental impact varies significantly by inference location. We use these regional factors:

RegionCIF (gCO₂/kWh)PUESite WUE (L/kWh)Source WUE (L/kWh)
Azure US East (Virginia)4501.180.301.50
Azure Sweden251.150.100.20
Google Global1301.11.080.80
AWS US East3501.150.201.30
China (general)5251.351.502.00
xAI Memphis5001.250.501.40

What Counts as "Water"

A model's water footprint has two parts: on-site water (evaporated by data-centre cooling towers) and off-site water (consumed generating the electricity the data centre uses). Our figures combine both where possible — but published sources differ in scope, and that matters when comparing providers.

Current best public on-site (site WUE) figures: AWS ~0.19 L/kWh, Microsoft ~0.49 L/kWh, Google fleet ~1.0–1.1 L/kWh, industry average ~1.8 L/kWh. Google's verified per-prompt figure covers on-site cooling water only, while our estimates for most other providers include off-site power-plant water — so Google models can look more water-efficient partly because of scope, not just engineering. We flag this on the relevant model pages.

Benchmark Figures vs Production Reality

A 2026 production-grounded study in Joule (Oviedo et al.) measured real fleet telemetry and found a median of just 0.31 Wh per query for frontier models (rising to ~3.9 Wh with test-time scaling) — and that benchmark-style estimates can overstate real fleet energy by 4–20×, because production serving batches requests and routes most traffic to efficient paths.

For that reason we no longer use the 2025 benchmark figures from Jegham et al. (for example 18–29 Wh for GPT-5, o3 and DeepSeek-R1) as headline numbers. They remain valid as upper bounds for a long reasoning prompt on 2025 hardware, and the relevant model pages still quote them.

Since the November 2026 review, text and code models use the CLEER basis instead: full-facility energy (accelerators, host servers, idle capacity and cooling) for one quick single question on current-generation hardware. These figures land close to the production medians above and are typically 5–15× lower than our earlier benchmark-derived estimates. CLEER also shows how much usage pattern matters: a typical multi-turn chat costs about 3.5× a single question, and a typical agentic coding session about 100×.

CLEER covers 51 models. Where it has a row for a model we use it directly. Where it does not, we extrapolate and say so on the model page: from the nearest sibling model; for open-weight models, from a fit of CLEER's benchmarks against parameters active per token (≈0.02 Wh + 0.005 Wh per billion); and for reasoning-only models, by multiplying the standard-chat figure by 5 (13 for the heaviest tiers). All of these are labelled estimated. Provider disclosures (Google's Gemini median, Mistral's life-cycle analysis) and the GPT-4o benchmark, which CLEER independently reproduces, are unchanged. Image, video, audio and music models are outside CLEER's scope and keep their earlier estimates.

Limitations

  • We measure inference only, not training or hardware manufacturing
  • Energy consumption varies 10–100× by prompt length and complexity
  • Provider disclosures may undercount overhead (cooling, networking, storage)
  • Self-hosted model data depends on your hardware and electricity source
  • Water and CO₂ figures are derived from energy using regional averages — actual values vary by time of day and season
  • We update data periodically; real-world efficiency improves continuously

Data last reviewed: November 2026

Calculate My Compute