Revolut moved 2,000 engineers onto a self-hosted Chinese model and says its AI coding bill fell by about 90%
On 3rd Nov'26 The Register reported a talk by David Abitbol, who leads AI operations at Revolut, in which he said the fintech cut its AI coding inference costs by roughly 90% in less than a year by making self-hosted open-weight models the default for about 2,000 engineers instead of commercial tools such as Claude Code. Revolut started on Z.ai's GLM-5.1 and now serves GLM-5.3 on 56 Nvidia B200 GPUs, at about $0.30 per million tokens by its own figures; because GLM is text only, a router sends image work to Gemini or GPT. The figures are Revolut's and the $0.30 rate has not been broken down. A day later Bloomberg reported that Apple had cut some engineers' daily AI allowance from $300 to $200 and is considering $100.
What it actually means
The 90% is the headline, but the useful part is what happened once the bill went flat. Revolut's first rollout of commercial tools came with strict budget limits, and Abitbol says that once the model ran on its own hardware, engineers stopped rationing tokens and spent them on debugging and Jira tickets, which is where Revolut puts its claimed 60% improvement in coding time. Apple is handling the same cost by lowering the cap, while Revolut took the meter away.
Few teams can stand up 56 B200s, and Revolut has not said what its $0.30 rate includes, so treat the saving as a ceiling. The smaller test is to run an open-weight model such as GLM through a regional cloud host on a week of ordinary tickets, and keep the frontier model for the work it gets wrong. Abitbol says the data risk with a Chinese model comes from calling its API in China rather than running the weights yourself, which your security team should check before you rely on it.
A per-engineer AI budget saves money, but Revolut found its engineers did more useful work once the meter was gone.
90%Revolut's claimed cut in AI coding inference costs
Source: The Register · Saturday, 7 November 2026