Revolut says it cut its AI coding inference costs by about 90% in under a year by making self-hosted models the default instead of commercial services like Claude Code, The Register reports. Ops and infrastructure lead David Abitbol spoke at the AI Engineer conference.
The numbers
- About 2,000 engineers, with a 60% improvement in coding time and 70% in lead time
- Self-hosting GLM-5.1 at about $0.30 per million tokens cost about a tenth of Codex or Claude Code
- It now serves Z.ai’s GLM-5.3 on 56 Nvidia B200 GPUs. GLM 5.1 is about 754B parameters (40B active) and needs about 380GB of GPU memory just for the weights
The interesting part
When tokens got cheap, engineers stopped rationing them. The biggest gains came from letting AI handle debugging and Jira and Confluence tickets, not from writing code. “When you have unlimited tokens, you’re free to do that.”
My take
That’s the Gartner point from yesterday, playing out at a real company. Open weights plus your own hardware means control and cost. Most of us won’t have 56 B200s, but the pattern scales down.
God Bless, and May the Source Be with You!
