Revolut moves its developers to open AI models on its own servers and now pays a tenth
The British fintech has largely dropped paid services such as Claude Code for its coding assistant and now runs Chinese open models on its own Nvidia chips.
Revolut has cut the cost of AI help for programming by about 90 percent by running open models itself instead of paying per use to commercial providers. David Abitbol, who leads operations and infrastructure at Revolut, said so at the AI Engineer conference in New York, The Register reports. StartupHub.ai published the same figures, based on the video of that talk.
According to The Register, Revolut earlier tried Gemini and, at the start of this year, Claude Code, but bills ran to between 1,000 and 10,000 dollars per engineer per month. The bank now runs models from the GLM family by China's Z.ai on 56 Nvidia B200 chips, at about 30 cents per million tokens, roughly a tenth of what the paid services cost. Some 2,000 engineers use the tools, and Revolut cites a 60 percent gain in coding time. StartupHub.ai notes that cloud, staffing and monitoring costs are not included in that token price and that the figures have not been independently verified.
The model handles text only, so questions about images and interface design are routed to other models. On the security of Chinese open models, Abitbol said according to The Register that the main risk lies in using them through services hosted in China, and much less in running them on one's own or regional servers.
Transparency: this edition was compiled by Claude, made by Anthropic.