CarbonWatch

CarbonWatch recognises 74 AI services. Each carries the type of use it serves and the region where it runs its inference — that region sets the carbon intensity applied, not the one you are in.

By kind of use

What the energy is spent on, rather than who profits from it.

Chat26 services
What people call a chatbot: a question, a written answer, produced token by token. It is the most common type, and the one whose estimate is the most direct — the bytes arriving are the answer being written.
Coding13 services
Completion in the editor, agents that write and run code, automated reviews. Usage is more continuous than in chat: the assistant answers every keystroke, not every question.
Search3 services
A question, an answer written from pages fetched and cited. The model reads far more than it writes, but only the answer crosses the network towards you.
Image, audio, video20 services
Nothing is counted in tokens here: what arrives is the media itself. A two-megabyte image is two megabytes of payload, not two megabytes of framing around some text.
Local models12 services
Ollama, LM Studio, llama.cpp and the others send nothing: the model runs here, and the electricity consumed is this socket's. It is the only case where the grid chosen in the settings applies.

The admission rule

A domain only enters the registry if it shares no IP address with obviously non-AI traffic. Detection works by address, never by hostname — nothing is decrypted. A shared address is therefore not an imprecision: it is foreign traffic billed as inference.

The 93 domains shipped were resolved and checked against a panel of Google, AWS, Cloudflare, Vercel and Akamai witnesses unrelated to AI. The current registry produces no collision, neither with those witnesses nor between two services.

What is deliberately absent, and why →

A file you can edit

The registry is a JSON file dropped into the application's data folder on first launch, and openable from its settings. Adding an entry is enough to have a private service recognised — an Azure OpenAI deployment, for instance, whose endpoint is specific to each customer.

Two optional fields refine a service whose type rate would be wrong: bytesPerToken, when its stream framing differs, and whPerKtoken, when the model served is not a frontier one. A translation engine or a style checker runs on a small, specialised model an order of magnitude cheaper: without that second field, widening the registry would amount to billing DeepL like GPT.

The registry, as data

The file is served as is, with no account and no key: it is the same list the application reads.

/registre.json

Every service

Service	Type	Inference region	Intensity	500-token answer
Claude	Chat	United States	380 g/kWh	0.38 g · 500 tokens
ChatGPT / OpenAI	Chat	United States	380 g/kWh	0.38 g · 500 tokens
Grok / xAI	Chat	United States	380 g/kWh	0.38 g · 500 tokens
Meta AI	Chat	United States	380 g/kWh	0.38 g · 500 tokens
Mistral / Le Chat	Chat	France	50 g/kWh	0.05 g · 500 tokens
DeepSeek	Chat	China	580 g/kWh	0.58 g · 500 tokens
Microsoft Copilot	Chat	United States	380 g/kWh	0.38 g · 500 tokens
Poe	Chat	United States	380 g/kWh	0.38 g · 500 tokens
Perplexity	Search	United States	380 g/kWh	0.48 g · 500 tokens
OpenRouter	Chat	World	480 g/kWh	0.48 g · 500 tokens
Groq	Chat	United States	380 g/kWh	0.38 g · 500 tokens
Together AI	Chat	United States	380 g/kWh	0.38 g · 500 tokens
Fireworks	Chat	United States	380 g/kWh	0.38 g · 500 tokens
Replicate	Image, audio, video	United States	380 g/kWh	1.52 g · 2 MB image
Cohere	Chat	United States	380 g/kWh	0.38 g · 500 tokens
Hugging Face Inference	Chat	World	480 g/kWh	0.48 g · 500 tokens
GitHub Copilot	Coding	United States	380 g/kWh	0.57 g · 500 tokens
Cursor	Coding	United States	380 g/kWh	0.57 g · 500 tokens
Windsurf / Codeium	Coding	United States	380 g/kWh	0.57 g · 500 tokens
Sourcegraph Cody	Coding	United States	380 g/kWh	0.57 g · 500 tokens
ElevenLabs	Image, audio, video	United States	380 g/kWh	1.52 g · 2 MB image
Midjourney	Image, audio, video	United States	380 g/kWh	1.52 g · 2 MB image
Suno	Image, audio, video	United States	380 g/kWh	1.52 g · 2 MB image
Runway	Image, audio, video	United States	380 g/kWh	1.52 g · 2 MB image
Goose (agent)	Coding	World	480 g/kWh	0.72 g · 500 tokens
Aider (agent)	Coding	World	480 g/kWh	0.72 g · 500 tokens
Ollama (local)	Local models	this machine	—	0.75 Wh/min · on this machine
LM Studio (local)	Local models	this machine	—	0.75 Wh/min · on this machine
llama.cpp (local)	Local models	this machine	—	0.75 Wh/min · on this machine
Jan (local)	Local models	this machine	—	0.75 Wh/min · on this machine
Sora	Image, audio, video	United States	380 g/kWh	1.52 g · 2 MB image
Luma / Dream Machine	Image, audio, video	United States	380 g/kWh	1.52 g · 2 MB image
Pika	Image, audio, video	United States	380 g/kWh	1.52 g · 2 MB image
Kling	Image, audio, video	China	580 g/kWh	2.32 g · 2 MB image
HeyGen	Image, audio, video	United States	380 g/kWh	1.52 g · 2 MB image
Synthesia	Image, audio, video	European Union	250 g/kWh	1 g · 2 MB image
Ideogram	Image, audio, video	United States	380 g/kWh	1.52 g · 2 MB image
Recraft	Image, audio, video	United States	380 g/kWh	1.52 g · 2 MB image
Leonardo	Image, audio, video	United States	380 g/kWh	1.52 g · 2 MB image
Stability AI	Image, audio, video	United States	380 g/kWh	1.52 g · 2 MB image
Black Forest Labs / Flux	Image, audio, video	European Union	250 g/kWh	1 g · 2 MB image
Civitai	Image, audio, video	United States	380 g/kWh	1.52 g · 2 MB image
Udio	Image, audio, video	United States	380 g/kWh	1.52 g · 2 MB image
Adobe Firefly	Image, audio, video	United States	380 g/kWh	1.52 g · 2 MB image
Krea	Image, audio, video	United States	380 g/kWh	1.52 g · 2 MB image
Qwen / Tongyi	Chat	China	580 g/kWh	0.58 g · 500 tokens
Kimi / Moonshot	Chat	China	580 g/kWh	0.58 g · 500 tokens
Zhipu / GLM	Chat	China	580 g/kWh	0.58 g · 500 tokens
Ernie / Baidu	Chat	China	580 g/kWh	0.58 g · 500 tokens
Doubao / ByteDance	Chat	China	580 g/kWh	0.58 g · 500 tokens
Z.ai	Chat	China	580 g/kWh	0.58 g · 500 tokens
Character.AI	Chat	United States	380 g/kWh	0.38 g · 500 tokens
Pi / Inflection	Chat	United States	380 g/kWh	0.38 g · 500 tokens
You.com	Search	United States	380 g/kWh	0.48 g · 500 tokens
Phind	Search	United States	380 g/kWh	0.48 g · 500 tokens
AWS Bedrock (US)	Chat	United States	380 g/kWh	0.38 g · 500 tokens
AWS Bedrock (UE)	Chat	European Union	250 g/kWh	0.25 g · 500 tokens
Continue	Coding	United States	380 g/kWh	0.57 g · 500 tokens
Tabnine	Coding	United States	380 g/kWh	0.57 g · 500 tokens
Zed AI	Coding	United States	380 g/kWh	0.57 g · 500 tokens
Bolt.new	Coding	United States	380 g/kWh	0.57 g · 500 tokens
Lovable	Coding	World	480 g/kWh	0.72 g · 500 tokens
Devin	Coding	United States	380 g/kWh	0.57 g · 500 tokens
Amp / Sourcegraph	Coding	United States	380 g/kWh	0.57 g · 500 tokens
DeepL	Chat	European Union	250 g/kWh	0.06 g · 500 tokens
Grammarly	Chat	United States	380 g/kWh	0.1 g · 500 tokens
vLLM (local)	Local models	this machine	—	0.75 Wh/min · on this machine
LocalAI (local)	Local models	this machine	—	0.75 Wh/min · on this machine
GPT4All (local)	Local models	this machine	—	0.75 Wh/min · on this machine
Msty (local)	Local models	this machine	—	0.75 Wh/min · on this machine
AnythingLLM (local)	Local models	this machine	—	0.75 Wh/min · on this machine
ComfyUI (local)	Local models	this machine	—	0.75 Wh/min · on this machine
KoboldCpp (local)	Local models	this machine	—	0.75 Wh/min · on this machine
RamaLama (local)	Local models	this machine	—	0.75 Wh/min · on this machine

See the global total →