Revolut moves 2,000 engineers to self-hosted GLM, cutting AI coding bill 90%

Revolut moved about 2,000 software engineers from Claude Code and Codex to a self hosted GLM model from Chinese AI lab Z.ai, David Abitbol, who oversees AI operations and infrastructure at Revolut, said in a talk at the AI Engineer conference in New York, Trending Topics reports. Revolut says its model bill is now about one tenth of what it previously paid Anthropic and OpenAI. An analysis by StartupHub.ai notes that the savings figure leaves out cloud, staff and monitoring costs.

The company runs the model, which has about 800 billion parameters, on roughly 56 Nvidia B200 chips. Revolut estimates a cost of about 30 cents per million tokens and no longer imposes token limits on engineers. In a sample of about 2,000 engineers, coding time fell 60% and the time from starting a task to merging it fell 70%.

Revolut uses the openly available model weights and hosts them on infrastructure in the United States rather than using Z.ai's inference service in China. Its coding agents run in a sandbox with command approval rules and require human oversight.