Revolut says it has cut AI inference costs by as much as 90% after replacing Claude workloads with self-hosted, open-weight models from China’s Z.ai. The company says the change has also allowed developers to use AI more freely without token limits.
Revolut Infrastructure Lead David Abitbol said the bank had initially placed strict budget limits on AI use because it was skeptical about the cost. Some heavy users were reportedly spending thousands of dollars a month on tokens without a comparable return, after the company had previously tested Google’s Gemini and then adopted Anthropic’s Claude.
Speaking at GTC Berlin, Abitbol said the self-hosted setup improved lead-time productivity by 70% and coding-time productivity by 60%. He also highlighted the privacy and auditability of Z.ai’s GLM models.
The article reports that GLM-5.1 brought inference costs to about $0.30 per million input and output tokens, roughly one-tenth the cost of comparable commercial services. Revolut can now manage its own inference capacity, but running the models requires substantial infrastructure. The company reportedly operates 56 Nvidia B200 GPUs, while the model needs around 380GB of GPU memory, before additional memory requirements for users’ context are considered.
Although self-hosting has reduced direct model usage costs, hardware and infrastructure bring their own expenses, including maintenance and electricity. Abitbol said open source is now a priority at Revolut, adding that its economics are favorable and its billing is predictable, without unexpected monthly charges.2
Sources: TechRadar.