Revolut has moved its AI coding inference off commercial services like Claude Code and onto a self-hosted GLM model, running on 56 Nvidia B200s and serving around 2,000 engineers. Costs reportedly fell by about 90 percent in less than a year. All figures come from a conference talk given by Revolut, not from an independent audit.
To me, this is the prompt-tax story told from the other side: once you have enough volume, you stop renting intelligence back by the token and start building it yourself. What's interesting is what stays outside. Image requests still go to third-party models, because GLM only does text. So Intelligence Ownership doesn't mean self-sufficiency. It means deciding deliberately which dependencies you keep.
And when a bank does the maths in public, the “too complex for us” argument loses a good deal of its air.