Revolut cuts AI coding costs tenfold by switching to self-hosted models

[Hattie35]

Revolut has found a much cheaper way to give its engineers AI coding tools. The Register reports that the British fintech cut its AI coding inference costs by about 90 percent in under a year after making self hosted models the default

David Abitbol, Revolut's operations and infrastructure lead, told the AI Engineer conference in New York that the company moved from Gemini in early 2025 and Claude Code in early 2026 to Z.ai's open weight GLM-5.1 and then GLM-5.3 models. They run on 56 Nvidia B200 GPUs, around seven DGX B200 systems

Self hosting costs about $0.30 per million tokens, roughly a tenth of what Codex or Claude Code cost the company, and some engineers had previously been spending between $1,000 and $10,000 a month. Around 2,000 engineers use the system, and Revolut claims 60 percent better coding time and 70 percent better lead time, with the biggest gains in debugging and handling Jira and Confluence tickets rather than writing code

The team targets under three seconds to the first token with a 256,000 token context and about 200 tokens per second, and found that tuning and NVFP4 quantisation mattered more than adding GPUs. GLM is text only, so vision tasks still go to Gemini or GPT, and Abitbol acknowledged concerns about using a Chinese open weight model. All the figures are company claims

This shows how quickly open weight models have caught up for everyday coding tasks, and how much money large companies can save at scale. It also raises questions about data security and model origins that each company has to weigh up. Expect more firms to try similar setups

Would you trust a self hosted open model for coding at work? And are big AI subscriptions becoming too expensive for companies?

British fintech Revolut slashes AI coding costs 10x by cutting out Claude and going self-hostedwww.theregister.com

[Glassy Viper]

Debugging and tickets benefiting most, rather than code writing, is interesting. AI helps most with the boring parts of the job. That matches what I have seen. Code generation gets the hype, but admin is where time goes. Admin work is where AI shines

[Dan_38]

Some engineers spending $10,000 a month on AI tools is wild. No wonder Revolut looked for alternatives

Costs like that cannot scale. Engineers love cheaper tools

[SašaJelenič]

Fifty six B200 GPUs is a big upfront investment. The savings only work at scale. Smaller companies would struggle to justify it. Upfront cost versus monthly bills