TL;DR

Revolut has moved AI coding for about 2,000 engineers onto self-hosted open-weight models from China’s Z.ai, cutting its inference bill by roughly 90% in under a year, the company’s David Abitbol told a New York conference in October, The Register reported on 3 November. These are the company’s own figures. In our view, a British fintech publicly choosing a Chinese model is notable in itself.

Abitbol, who leads operations and infrastructure at the fintech, described a company that was wary of AI costs from the start. Staff were spending “$1,000 a month, $10,000 a month with not so much ROI”, he said, so the first rollout of commercial tools (Claude Code followed Gemini) came with hard budgets. Token limits then led the firm to look at running models itself.

The Register puts GLM 5.1 at $0.30 per million tokens, which worked out at about a tenth of the price of Codex or Claude Code. Revolut has since moved to GLM-5.3, running on 56 Nvidia B200 GPUs. It also reports coding time improving by 60% and lead time by 70%, without saying what those measures cover in the report.

Where the gains came from

The biggest gains were not in writing code. They came, Abitbol said, from handing debugging and ticket admin to the models, because “you no longer have to save your tokens”. On infrastructure, tuning and quantisation mattered more than extra GPUs, and he called very long context windows “a scam”, arguing 256,000 tokens is enough.

GLM has not replaced everything. It handles only text, so a router sends vision work to Gemini or GPT Luna.

The question for UK firms

Abitbol accepts that Chinese open-weight models raise privacy and security concerns, but says the main risk lies in reaching them through APIs hosted in China, and that running them on-premises or on regional cloud removes the worry. The Register reports no independent assessment of that view. We have reported separately on how far GLM-5.3 got in simulated cyberattacks, which is a different question from data leakage but one a risk committee would ask too.

In our view, the economics deserve attention from UK technology leaders, with two cautions. Fifty-six GPUs and a team able to tune them are a large commitment, so a firm with a few dozen developers should not assume the same saving. And flat costs depend on how much the hardware gets used, a figure the report does not give.