Revolut cut AI coding costs 90% by self-hosting GLM for 2,000 engineers, not by switching providers

Revolut self-hosts GLM 4.5 on its own infrastructure, running it across 56 Nvidia B200 GPUs, with the OpenAI-compatible setup plugging into tools like OpenCode and Codex. Image-generation tasks still get routed to Gemini or similar models elsewhere, so the self-hosted path covers text and coding work specifically. The effective cost per million tokens comes out to about $0.30, roughly a tenth of what Revolut had been paying Anthropic and OpenAI for the same workloads, and that comparison is what produces the 90% figure. That math excludes cloud infrastructure, staffing, and observability costs tied to running the self-hosted setup, and getting there took real engineering work rather than a flip of a switch: Revolut’s team has said out-of-the-box inference performance was poor, and raised throughput 88% while demand grew 76% without adding GPUs, through tuning context limits, quantization, and parallelism settings specific to GLM’s architecture.

On the productivity side, Revolut reports coding time falling about 60% and lead time, the span from picking up a task to merging it, falling about 70% across the roughly 2,000-engineer sample after the switch. Time to first token dropped from more than 10 seconds to 1.6 seconds, and the system now handles more than 7 requests per second, up from 3 before the retuning. Removing hard token limits appears to be a real driver behind some of that lead-time improvement specifically, since engineers no longer had to ration usage and reportedly started automating smaller non-coding tasks, like inbox triage, that they’d previously just done by hand.

The catch is that this result assumes a company with 56 B200 GPUs sitting idle to dedicate, an engineering team capable of the tuning work involved, and enough scale to justify building the infrastructure in the first place. A smaller team trying to replicate Revolut’s numbers by simply pointing their agents at a hosted GLM API, without the self-hosting and tuning behind it, would likely see a smaller gap between the old and new costs.