Revolut says self-hosted models cut its AI coding costs by about 90%
90% cut in AI coding costs, per Revolut
Revolut says it cut its AI coding inference costs by about 90 percent in under a year by making self-hosted open-weight models its default, The Register reported on November 3, 2026. It runs Z.ai’s GLM model on 56 Nvidia B200 GPUs for about 2,000 engineers and routes work the model can’t handle, such as image tasks, to Gemini or GPT. David Abitbol, Revolut’s ops and infrastructure lead, told an October conference that the biggest gains came from debugging and Jira and Confluence tickets, not faster coding. The figures are Revolut’s own, with no total spend or outside check.
When you have unlimited tokens, you're free to do that.
Alan's takeThis is fantastic! The right models, housed in the right place, will unlock a ton of value. As of this writing, Artificial Analysis tracks more than 700 models. We’ll eventually need thousands of purpose-specific ones.