Estimator methodology

Version 2026.06 · updated 2026-06-12. Every number below is versioned in the open and carries a citation — when a provider publishes a new disclosure, one row changes.

The estimation chain

energy (Wh)     = (tokens / 1000) × Wh-per-1k-tokens(model class) × PUE
emissions (gCO₂e) = energy (kWh) × grid intensity of region (gCO₂/kWh)

Results are always shown as a range: the honest uncertainty on per-token energy is ±3–5×, so a single number would be false precision. When you enter queries instead of tokens, we assume 500 output tokens per query (a typical chat answer).

1 · Energy per token

Anchored on 2025 provider disclosures — Google's median Gemini text prompt (0.24 Wh / 0.03 gCO₂e, including idle machines, CPU/RAM and datacenter overhead), Epoch AI's ~0.3 Wh per average ChatGPT query, and Mistral's LCA (~1.14 gCO₂e per 400-token answer). You pick a model class, not an exact model:

Class	Examples	Wh / 1k output tokens
Light	Haiku, GPT-4o-mini, Gemini Flash	0.12–0.25
Standard	GPT-4o, Sonnet, Gemini Pro	0.4–0.8
Frontier / reasoning	Opus, o-series, extended thinking	1.5–6

Frontier/reasoning models include hidden thinking tokens. The often quoted “3 Wh per query” figure predates provider disclosures and measured much less efficient 2023-era serving — we treat it as obsolete.

2 · Grid intensity

Your cloud region maps to a physical grid. We use location-based annual intensities (Ember) — the physical reality of the grid the datacenter draws from. Providers usually advertise market-based figures (renewable PPAs and certificates), which can show “0 g” on a coal-heavy grid; we note them but do not use them. A live ENTSO-E layer for EU regions is on the roadmap via our agents service.

Region	Grid	gCO₂/kWh	Cloud regions
Europe (Stockholm)	Sweden (hydro + nuclear)	25	AWS eu-north-1 · Azure Sweden Central
Canada (Montréal)	Canada — Québec (hydro)	30	AWS ca-central-1 · GCP northamerica-northeast1
Europe (Paris)	France (nuclear)	55	AWS eu-west-3 · Azure France Central
Europe (London)	United Kingdom	220	AWS eu-west-2 · Azure UK South
Europe (Amsterdam)	Netherlands	270	Azure West Europe · GCP europe-west4
US West (Oregon)	United States — Oregon	285	AWS us-west-2 · GCP us-west1
Europe (Ireland)	Ireland	290	AWS eu-west-1 · Azure North Europe
US Central (Iowa)	United States — Iowa	340	GCP us-central1 · Azure Central US
Europe (Frankfurt)	Germany	350	AWS eu-central-1 · GCP europe-west3
US East (N. Virginia)	United States — Virginia	370	AWS us-east-1 · Azure East US · GCP us-east4
Asia (Tokyo)	Japan	450	AWS ap-northeast-1 · Azure Japan East
Asia (Singapore)	Singapore	470	AWS ap-southeast-1 · GCP asia-southeast1
Australia (Sydney)	Australia — NSW	510	AWS ap-southeast-2 · Azure Australia East
Asia (Mumbai)	India	630	AWS ap-south-1 · GCP asia-south1

3 · Vendor factors

The vendor comparison applies a multiplier to the model-class baseline (managed, well-batched serving at hyperscale PUE ~1.1) and maps each vendor to the regions it actually serves from. These are honest estimates, not disclosures — except Google, the only provider that has published a per-prompt figure.

Vendor	Factor	Why
OpenAI	×1–1.2	Serves from Microsoft Azure, predominantly US — typically US East (Virginia).
Anthropic	×1–1.2	Serves from AWS and Google Cloud, US-heavy.
Google Gemini	×0.8–1.1	TPU serving on GCP; the only provider with a published per-prompt figure (0.24 Wh median).
Self-hosted (open weights)	×2–4	Own GPUs: lower batching and utilization than hyperscale serving, usually higher PUE.

4 · Overheads

PUE (Power Usage Effectiveness) — cooling and facility overhead on top of IT load. Default 1.2; adjustable 1.1–1.6 (hyperscalers report ~1.1, the global average is ~1.55).
Embodied carbon — chip manufacturing, servers and buildings; optional +30% uplift, a common LCA approximation.
Training is not amortized into per-query figures — it is a separate, one-off cost that would add false precision per query. We call it out instead of hiding it in the number.

Sources

Try it on your own workload — open the AI Footprint Estimator.

Research & Business plans · machine-readable export

↓ Download methodology.json

Versioned constants, sources and grid intensities as JSON. Upgrade to Research to unlock.