Revolut says it cut AI coding inference costs by roughly 90% by making self-hosted models its default, while deploying AI coding tools to about 2,000 engineers. Its experience offers a useful look at what it takes to make large-scale, self-hosted AI work beyond a promising demo.
Watch Desk analysis
What happened
Speaking at the AI Engineer conference in New York last month, Revolut infrastructure lead David Abitbol said the company moved from commercial coding services to open-weight models, first GLM-5.1 and then GLM-5.3. The Register reports that Revolut now serves GLM-5.3 on 56 Nvidia B200 GPUs.
Abitbol said the rollout of AI coding tools was associated with a 60% improvement in coding time and a 70% improvement in lead time. He also said self-hosting GLM-5.1 cost about $0.30 per million input and output tokens, roughly a tenth of using Codex or Claude Code. These are Revolut’s reported results, not a like-for-like independent benchmark.
The company’s setup is more than buying a rack of GPUs and hoping for the best. Abitbol pointed to model and context-length tuning, NVFP4 quantisation, and separating prefill from decode. He said Revolut targets about 256,000 tokens of context rather than using a one-million-token window, and reported under three seconds to first token and generation speeds around 200 tokens per second for its engineers.
GLM is text-only, so Revolut routes vision queries to models such as Gemini or GPT Luna instead of asking one model to do everything.
Why it matters
The reported savings came alongside a change in how engineers could use AI: Abitvol said cheaper tokens meant less need to ration usage, including for debugging and work involving Jira and Confluence. That makes the case about more than swapping a model licence for a GPU bill. Self-hosting changes the economics, but brings infrastructure, optimisation and model-routing work with it.
Our read
This is a substantial, practical account of an organisation taking AI coding in-house, not proof that every company can reproduce the results. The interesting lesson is that hardware was only part of the recipe; the serving setup and the decision not to chase the largest possible context window mattered too. Anyone considering a similar move should compare total operating costs and task performance, not token prices alone.
What to watch
- Whether Revolut shares more detail on how it measures cost and productivity changes.
- How its self-hosted setup performs as usage and model requirements change.
- Whether the split between self-hosted text models and external vision models changes over time.
Discussion spark: Would you trust a self-hosted open-weight model for most coding work if it cut costs sharply, or are the extra infrastructure and model-routing demands a false economy?
Sources and evidence
- British fintech Revolut slashes AI coding costs 10x by cutting out Claude and going self-hosted (3 November 2026, 20:07 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.