Sustainability overview

Our infrastructure runs on data centers powered entirely by renewable energy sources, hosted by Scaleway in France. Their facilities are independently certified to use 100% renewable electricity, verified through regular external audits.

PUE (Power Usage Effectiveness) measures how efficiently a data center uses its electricity: a PUE of 1.0 would mean every watt powers the IT equipment itself. Our infrastructure operates at a PUE of 1.25, compared to the industry average of 1.55, meaning roughly 19.4% less electricity is lost to cooling and overhead.

WUE (Water Usage Effectiveness) measures the amount of water a data center consumes per unit of computing it delivers. Our hosting partner operates at a WUE of 0.25 L/kWh against an industry average of 1.8 L/kWh, an 86.1% reduction. This matters because traditional data centers consume large volumes of water for cooling, often in regions that are already water-stressed.

Our hosting is sustainable, but we can't yet verify that the underlying foundation models we use were trained with renewable energy. We're actively exploring partnerships with renewable energy providers specifically for AI model training, and we'll publish progress on this page as it lands.

We track energy consumption at the model and infrastructure level, then attribute it back to your usage. The result is per-conversation, per-token, and per-account energy and carbon metrics, available through your dashboard and exportable for your sustainability reporting.

Yes. Our methodology is published, and the measurements are auditor-ready. You can export totals for any time window, broken down by Scope 2 emissions where relevant, and drop them straight into your annual sustainability filing.

The data centers we run on maintain ISO 50001 (energy management) and ISO 27001 (information security) certifications, audited regularly by independent third parties.

We deliberately route each task to the smallest model that can answer it well. Lower-parameter models cut computational requirements by roughly 40–50% on the tasks they handle. Specialised lightweight agents handle simpler workflows, and prompt optimisations cut token usage by a further 20–30%.