British fintech Revolut slashes AI coding costs 10x by going self-hosted

Revolut says it cut AI coding inference costs by about 90 percent in under a year by making self-hosted models its default, rather than relying mainly on services such as Claude Code. The fintech rolled AI tools out to roughly 2,000 engineers and reported coding times improved by 60 percent and lead times by 70 percent. It first imposed strict token budgets after seeing some engineers spend thousands of dollars monthly with limited returns; running the open-weights GLM model cost about one-tenth as much as commercial alternatives. With cheaper tokens, engineers could use AI more freely, without rationing it for only the hardest tasks.

The biggest productivity gains, Revolut’s infrastructure lead said, came not simply from generating code but from using AI for debugging and handling Jira and Confluence tickets. The company now serves GLM-5.3 on 56 Nvidia B200 GPUs, tuning quantization, model settings, context lengths, and the division of prefill and decoding work to improve performance. It targets a roughly 256,000-token context, sub-three-second response startup, and around 200 generated tokens per second, finding million-token contexts unnecessary. GLM does not replace every model: it is text-only, so a router sends visual tasks to systems such as Gemini or GPT. Revolut favors self-hosting for predictable costs and says regional or on-premises deployment addresses its main security concern about Chinese models: sending data through APIs hosted in China.