Revolut cuts AI cost to $0.30 with in-house GLM-5.1

Revolut's shift began with a budgeting problem. After adopting Google's Gemini in early 2025 and Anthropic's Claude Code in early 2026, the company saw some engineers consume $1,000 or even $10,000 worth of AI per month. Setting spending caps led engineers to ration usage, reserving AI for only the hardest tasks and reducing overall use. Moving to self-hosting was a response to that trade-off, and Revolut chose Z.ai's open-weight GLM series, first GLM-5.1 and later GLM-5.3.

The economics depend on more than raw GPU count. Running models on owned hardware required substantial compute capacity, which Revolut met with 56 Nvidia B200 GPUs. Avitbol attributed about 80% of the performance gains to adjustments in how the AI is configured, including controlling how much information it references at once and quantization, which reduces data precision to save memory. That suggests the cost advantage is as much an operations and tuning result as a hardware one.

Revolut has not abandoned commercial services. It continues to use Gemini where image understanding is needed, and Avitbol said self-hosted AI is now the default, with high-performance commercial services reserved for specific situations. One stated benefit is predictability: self-hosting makes operating costs easier to forecast and reduces the risk of a surprisingly large bill at month's end.