What this actually costs to run

1. This website is deliberately light. Plain server-rendered pages, one stylesheet, no JavaScript, no tracking scripts, no third-party requests, no video. The font is a subset rather than a full family, and the artwork is WebP with a smaller variant served to phones. A page here is a few tens of kilobytes, which is a rounding error against a typical marketing site and stays that way because the pre-deploy check refuses anything loaded from another host.
2. The application is an ordinary web stack. PHP and MariaDB, which runs on modest hardware and does not need a fleet of anything. There is no client-side model, so nothing is downloaded to a laptop or a phone to run locally.
3. The language model calls are the real cost, and we do not currently measure them. Every translated turn and every fidelity score is an inference request on somebody else's infrastructure. We do not have a figure for that and we are not going to invent one or buy an offset to cover a number we have not calculated. Measuring it properly is a fair question for the academic work being arranged.
4. Less travel is the offsetting effect that actually exists. A conversation that happens today is a second appointment that does not, and often a journey the person does not have to make. That is a real reduction and it is also unmeasured, so treat it as a reason rather than a claim.

What we will not do

1. Buy an offset to cover a number we have not calculated. An offset against an unmeasured figure is a purchase, not a reduction.
2. Publish a carbon figure we cannot show the working for. If it appears here it will come with the method.
3. Claim the travel saving as a number. It is real and it is unmeasured, so it stays a reason and not a statistic.

Measuring the inference cost properly is a fair question for the academic work being arranged. If your procurement needs an environmental figure we cannot yet give you, ask and you will get a straight answer about what is and is not known.