Most conversations about AI and the environment focus on the enormous energy cost of large model training. That's a fair concern. But there's a related question that gets less attention: what does it cost, environmentally, to runAI in production — every day, for every user?
At Vitreon Legal, we've made deliberate architectural choices that result in a minimal environmental footprint. Not as a marketing position, but because building lean is how we believe AI should be engineered.
What our infrastructure actually looks like
Our production stack runs on a Mac Studio and a single RTX 3070 consumer GPU. That's the compute that powers our embedding pipeline, our reranker, and our AI research assistant — serving real users across four jurisdictions every day.
No A100 cluster. No rented GPU racks. No cloud instances scaling elastically to hundreds of cores on demand. We've built an efficient, purpose-built retrieval pipeline that punches well above its weight class — our GaRAGe retrieval scores 36% above published state of the art — without needing massive compute to do it.
Efficiency-first AI architecture has a direct environmental benefit: less compute means less energy. That's not incidental — it's a consequence of refusing to solve hard problems by throwing more hardware at them. When your infrastructure budget is limited, you're forced to think carefully. That constraint has made us better engineers and produced a smaller carbon footprint.
Remote and async-first means zero office footprint
Our team has no office. There are no commutes, no building to heat or cool, no cafeteria, no infrastructure overhead of a physical workspace.
Async-first work is our default operating model — not a pandemic-era holdover. It means no mandatory travel for meetings, no relocations, no flights for offsites. The environmental consequence is real: our organisational footprint is essentially the footprint of the people themselves, wherever they already live.
What we're not claiming
We don't publish formal carbon metrics. We're not audited by an environmental third party. We're a startup, not a regulated entity, and we won't pretend otherwise.
What we can say plainly: the structural choices we've made — lean inference architecture, no cloud GPU clusters, no office, no mandatory travel — have genuine environmental consequences that are worth stating directly.
As we scale
As Vitreon grows, we're committed to maintaining this philosophy: add compute only when efficiency improvements have been exhausted first. Prefer smaller models where accuracy holds. Avoid the arms-race approach to AI infrastructure that has made “AI” synonymous with enormous energy consumption in public perception.
Legal research AI doesn't need to cost the planet. We've demonstrated that a retrieval pipeline purpose-built for legal text can outperform much more expensive infrastructure. That efficiency is both a competitive advantage and an environmental one.