The renewable AI assistant

AI that gives back more energy than it takes.

Training a large language model can burn through the yearly electricity of hundreds of homes. We don't pretend otherwise. GreenAI runs on renewable power, routes every question to the leanest model that can answer it, and funds new renewable capacity until the balance tips positive.

Start chatting free See our energy figures

No carbon guilt bolted on afterwards — energy accounting is built into how the product works.

100%Renewable-matched inference

2.6×Energy funded vs. energy used

<1.2Target datacentre PUE

1clickTo read your own footprint

The honest bit

Large language models are enormously power-hungry. Ignoring that isn't an option.

A single frontier model can take tens of gigawatt-hours to train — comparable to the annual electricity use of a small town — and every message you send afterwards draws a little more. Most AI companies quietly buy offsets and move on. We think the honest response is to measure it, minimise it, and then over-correct until the numbers point the other way.

Training is the big upfront debt. It happens once, but it's huge. We treat it as a loan to repay in renewable energy, not a cost to hide.
Inference is the daily drip. Small per message, but it adds up across millions of chats — so the model you talk to matters enormously.
Grids are still mixed. "Renewable" electricity often means certificates, not clean electrons on the wire. We fund genuinely new capacity, not just paperwork.

Our approach

Three commitments, in the order they matter.

Use less

The greenest kilowatt-hour is the one you never spend. We default to the smallest capable model, quantise aggressively, batch requests, and cache. Efficiency comes first — before any offset is even counted.

Run on renewables

Inference runs in datacentres powered by wind, solar and hydro, matched to real generation hour-by-hour where we can, not just annual averages. We publish where our compute physically runs.

Give back more

A fixed share of every subscription funds new renewable projects — community solar, wind, storage — sized to repay training and then keep going. The goal is net-positive, not net-zero.

The models we run

We pick the most efficient model that can do the job — and tell you which one answered.

Bigger isn't greener. Most questions are handled beautifully by compact, open models running quantised on modest hardware. We route heavier reasoning to larger models only when a task genuinely needs it.

Mistral Small 3

24B · everyday

Fast, capable general chat at a fraction of frontier-model energy. Our default for most conversations.

Llama 3.2

1B–8B · lightweight

Tiny variants run on very little power for quick lookups, drafting and summarising.

Qwen2.5

7B · multilingual

Strong coding and non-English performance without reaching for a giant model.

Gemma 2

2B–9B · efficient

Google's compact open models — excellent quality per watt for structured tasks.

Phi-4-mini

3.8B · reasoning

Punches well above its size on maths and logic, so we rarely escalate.

Llama 3.3 70B

70B · heavy lifting

For genuinely hard reasoning and long context. Used only on escalation, so its bigger draw stays the exception, not the default.

Qwen2.5 72B

72B · deep coding

Reserved for complex code and analysis that the compact models can't finish cleanly on their own.

DeepSeek-V3

671B MoE · sparse

A mixture-of-experts model that only activates ~37B parameters per token, so it delivers frontier quality at a fraction of a dense model's energy.

Llama 3.1 405B

405B · frontier

Our top tier for the hardest problems. Powerful but power-hungry, so the router reaches for it rarely — and always tells you when it did.

Kimi K2

1T MoE · 32B active

A trillion-parameter mixture-of-experts model that lights up just 32B parameters per token — frontier-scale reasoning without frontier-scale energy per query.

Smart routing

how it fits together

A lightweight router reads your request and sends it to the leanest model likely to nail it — escalating to the heavier models above only when confidence is low. You always see which model replied.

Our green credentials

The full accounting — including the parts we haven't solved yet.

Borrowed from the honest tradition of green infrastructure: state what's genuinely clean, and be candid about what still isn't.

Where the compute runs

Inference is placed in datacentres on renewable power-purchase agreements — onshore and offshore wind, solar and hydro — chosen for a low PUE (target under 1.2) so less energy is lost to cooling and overhead.

Efficiency by default

Smallest capable model, 4-bit quantisation, request batching and response caching. Every optimisation that cuts a watt ships before we count a single offset.

New capacity, not just certificates

We fund additional renewable generation and storage that wouldn't otherwise exist, sized to repay our training energy and then run net-positive. Certificates alone don't count.

Published figures

An annual energy report, and a per-conversation estimate you can open any time. If you can't see the numbers, it isn't accountability.

What we haven't solved

Grid mixes still aren't 100% renewable everywhere our compute lands. Hardware carries embodied carbon from manufacturing. We don't control upstream chip fabrication. We'd rather name these than airbrush them.

How you can check us

Our methodology and offset partners will be listed publicly. Hold us to the net-positive claim — that's the whole point of publishing it.

Ask something. Watch the energy stay in balance.

Start a conversation for free. Every chat shows which model answered and roughly what it cost in energy — with renewable investment already covering more than the draw.