How Revolut Slashed AI Coding Costs by 90% by Ditching Claude for Self-Hosted GLM

Revolut deployed Z.ai's open-source GLM model on a compact private cluster to service roughly two thousand software developers. The transition started because individual heavy users were burning through eye-watering monthly budgets under early deployments of Gemini and Claude Code, forcing management to impose restrictive token rationing. Seeking an escape from commercial lock-in, the engineering leadership rolled out GLM-5.1 before upgrading production workloads to GLM-5.3.

The entire heavy-lifting operation fits neatly onto just fifty-six Nvidia B200 accelerators packed inside seven standard DGX B200 systems. The active model weights consume roughly 380 gigabytes of video memory, leaving plenty of breathing room for user context windows. Performance was squeezed out not by buying endless hardware trucks, but through aggressive inference hyperparameter tuning like NVFP4 quantization and separate execution pools for prefill and decode stages.

Discarding the industry-wide obsession with million-token context windows proved critical to keeping hardware bills grounded. By capping active context at 256,000 tokens, the infrastructure easily holds around twenty-five thousand lines of code while delivering time-to-first-token responses under three seconds. Offloading routine repository queries and ticket management to an unlimited local assistant supercharged team productivity by seventy percent without triggering unexpected finance department panic attacks.

Of course, trusting Chinese model weights routed through local infrastructure comes with its own unique flavor of corporate anxiety and proprietary evaluation blind spots. Yet as open weights narrow the capability gap with proprietary monoliths, the temptation to bring workloads back on-premise turns every cloud invoice into an expensive relic of a wasteful past.

Source: The Register