Revolut says self-hosted AI cut inference costs by up to 90%

Revolut says moving some AI coding workloads from Anthropic’s Claude to self-hosted open-weight models from China’s Z.ai cut inference costs by as much as 90% in less than a year. The London-headquartered digital bank has not abandoned AI; it says developers can use the models more freely without token limits.

Revolut Infrastructure Lead David Abitbol said the company was “highly skeptical about the cost of AI” and initially imposed strict budget limits. Speaking at GTC Berlin, he said the self-hosted option improved lead-time productivity by 70% and coding-time productivity by 60%. He also cited the Z.ai GLM models’ data privacy and auditability.

The article reports that GLM-5.1 brought inference costs to around $0.30 per million input and output tokens, about one-tenth the price of comparable commercial services. Revolut now operates its own inference infrastructure, but self-hosting also brings hardware and operating expenses. The report says the bank has 56 Nvidia B200 GPUs. The model alone needs about 380GB of GPU memory, while context retention is estimated to require another 6 to 26GB per user.

Abitbol said open source is now a priority at Revolut and that its economics are favorable, with flat billing and no unexpected end-of-month bill. He described frontier models as the exception. The figures and comments are attributed to Abitbol and the report; the article does not quantify Revolut’s total infrastructure costs.