Revolut says self-hosted GLM cut its AI coding bill by 90%
The fintech runs the open-weight model for about 2,000 engineers on 56 Nvidia B200 GPUs, while Claude, Codex and vision models remain exceptions.
Revolut has made a self-hosted version of Z.ai’s GLM its default model for AI-assisted software work, according to David Abitbol, who leads AI operations and infrastructure at the fintech. The deployment serves roughly 2,000 engineers using 56 Nvidia B200 GPUs.
Abitbol said Revolut’s blended inference cost is about $0.30 per million tokens and that its bill is one-tenth of what it previously spent on Anthropic and OpenAI services. The company also reported a 60% improvement in coding time and a 70% improvement in engineering lead time after adopting GLM.
The practical change was not only cheaper code generation. Lower costs let Revolut stop rationing tokens, so engineers could use AI for debugging and routine work in Jira and Confluence instead of reserving it for difficult tickets. Abitbol said open-weight models are now the priority, while frontier systems such as Claude and Codex remain available as exceptions.
GLM has not replaced every external model. Revolut routes image-related requests to systems such as Gemini or GPT because its main GLM deployment is text-only. The claimed 90% reduction describes Revolut’s inference bill; the presentation does not provide a full comparison that includes the capital and staffing costs of operating the GPU fleet.
