The environmental cost of AI (and what gets done about it)

Most AI companies talk about capability and almost never about cost: not the invoice, the environmental one. It's not a minor footnote: training and running large models draws real electricity and, at data-center scale, real water for cooling. Ignoring that isn't neutral, it's a choice.

Where the cost sits

Training a frontier model is the headline number, but it's a one-time cost spread across millions of users. The bigger, less visible cost is inference: the same oversized model answering the same simple question millions of times a day, when a model a tenth the size would have done the job just as well. That's the part almost every business controls.

The default that makes it worse

The industry habit is to reach for the biggest, newest model regardless of the task. It's impressive in a demo and nobody gets blamed for over-provisioning. For a support agent answering questions from a fixed knowledge base, or a workflow classifying support tickets, that's not better performance. It's wasted compute wearing a performance costume.

What a smaller-by-default policy looks like in a real build

Start with the smallest model that clears the eval set for the task, and only move up when it genuinely can't hold the bar, not preemptively. Cache and reuse answers where the input repeats instead of re-generating them. Route only the parts of a workflow that need a large model's judgment there, and let smaller, cheaper models handle the rest. None of this is exotic; it's just treating compute as a cost worth managing, the way any other engineering resource is.

Why this is a mission, not a marketing line

Building AI responsibly means being honest about what it costs, both to the client's budget and to the planet running the servers, and defaulting to the option that's lighter on both wherever the job allows it. That standard doesn't get relaxed because a bigger model would look more impressive in a pitch.