Revolut says it cut AI coding spend by nearly 90% using open models on its own infrastructure

Revolut has made open models hosted on its own infrastructure the default option for its roughly 2,000 engineers. According to The Register, David Abitbol, head of operations and infrastructure, told the AI Engineer conference in New York that the move brought inference costs down by around 90% in less than a year.

Before that, the company had been using Gemini and Claude Code. Abitbol said some engineers were spending between US$1,000 and US$10,000 a month with no clear return. It now runs GLM-5.3 on 56 Nvidia B200 GPUs. In his view, the biggest performance gains came from tuning the deployment, with techniques like quantization and context limits, rather than from buying more hardware.

There are caveats. GLM only handles text, so a router sends queries that include images to Gemini or GPT. Abitbol also reported productivity gains, but those are the company's own numbers, with no independent verification.

Why it matters if you build with AI: not every team has 56 GPUs or someone to run them, so this recipe can't be copied as is. What you can take away is the question: measure how much you spend per task and look at which ones could go to a cheaper model through a router.