Revolut, a British neobank with Ukrainian roots, has moved nearly 2,000 of its engineers to a self-hosted, open-source model developed in China. The move not only significantly reduced the company’s costs but also fundamentally changed the way developers approach using artificial intelligence.
Revolut, a British neobank with Ukrainian roots, has moved nearly 2,000 of its engineers to a self-hosted, open-source model developed in China. The move not only significantly reduced the company’s costs but also fundamentally changed the way developers approach using artificial intelligence.
As David Abitbol, Head of AI Operations and Infrastructure at Revolut, told the AI Engineer conference in New York, the Chinese GLM 5.3 model from the Beijing-based company Z.ai has replaced Anthropic Claude Code and OpenAI Codex as the default tool.
Thanks to this transition, the neobank now spends only a tenth of what it previously paid to American AI laboratories, writes Cybernews.
Its own infrastructure allowed the company to get rid of unpredictable bills from third-party providers and give engineers complete freedom to use AI, as previously they were forced to save expensive tokens only for the most complex coding tasks.
Key indicators and integration parameters:
GLM 5.3 model deployed on 56 Nvidia B200 GPUs serving about 2,000 engineers;
the weighted average price is approximately $0.30 per million tokens;
The time spent writing code directly improved by approximately 60%, and the lead time (the time from the formation of requirements to the release of a feature to the user) was reduced by 70%.
As Abitbol explains, delays in development often arise from related tasks — working with email, Jira, and Confluence. Cheap access to tokens has allowed specialists to delegate a much wider range of routine processes to artificial intelligence.
A Revolut spokesperson acknowledges that using Chinese models carries “significant stigma,” especially for a fintech company regulated under EU law. However, he emphasizes that Revolut only uses publicly available model weights, runs them exclusively on US infrastructure, and always leaves a human in the decision-making chain.
