Revolut Cuts AI Coding Costs ~90% With Self-Hosted GLM on 56 B200s

Revolut's ops and infrastructure lead told The Register the London-based fintech cut its AI coding inference bill by roughly 90 percent in under a year by moving around 2,000 engineers off Claude Code and OpenAI's Codex onto a self-hosted deployment of Z.ai's GLM-5.3, served on 56 Nvidia B200 GPUs across seven eight-GPU DGX systems.

The arithmetic is blunt.

Self-hosted inference lands at "$0.30 per million input/output tokens," according to David Abitbol, Revolut's ops and infrastructure lead, against commercial seats that had been running engineers $1,000 to $10,000 a month. "We were seeing engineers spend $1,000 a month, $10,000 a month with not so much ROI," he said. Revolut credits the switch with a "60 percent improvement in coding time and a 70 percent improvement in lead time" after rollout.

The model itself, GLM-5.3, carries 754 billion total parameters with 40 billion activated per token, and needs about 380 GB of GPU memory for weights, with another 6 to 26 GB per user for context. The performance compression lives in the stack around it: NVFP4 quantization, disaggregated prefill and decode, hyperparameter tuning, and strict context-length caps. 256,000 tokens is Revolut's recommended ceiling, not a million. "Long context is a scam. You really need to be mindful of that," Abitbol said. He holds his team to a sub-three-second time to first token: "As soon as you exceed 5 seconds, my Slack inbox starts to blow up." Throughput sits near 200 tokens per second, 800 on the evening shift.

"Open source is now a priority at Revolut, and our economics are favorable," Abitbol told The Register. "The billing is flat, we don't have any bill shock at the end of the month, and the frontier is the exception." It is a sharper data point inside the 464 AI infrastructure stories our tracker has logged in the last 90 days: large users pushing the inference cost base down by running the weights themselves.