Revolut moves to self-hosted GLM

Revolut migrated approximately 2,000 engineers to self-hosted open-source models, reducing AI coding costs by 90% compared to Anthropic and OpenAI APIs while improving engineering lead time by 70%. The company deployed 56 Nvidia B200s at 30 cents per million tokens, with engineers only restricting frontier models to exceptions once token rationing was removed. The shift suggests open-weight models may become the industry default, posing significant pricing challenges for AI labs preparing for public offerings.