Revolut's Self-Hosted AI Coding Switch Cut Its Bill by 90 Percent

The 90 Percent Number Everyone Is Quoting

Revolut says its self-hosted AI coding switch cut inference costs by roughly 90 percent in under a year. The British fintech moved about 2,000 software engineers off Anthropic's Claude Code and OpenAI's Codex as their default coding tools, and onto GLM, an open-weight model from Chinese AI lab Z.ai, which Revolut hosts itself. The numbers come from David Abitbol, who runs AI operations and infrastructure at Revolut, in a talk at the AI Engineer conference in New York.

The setup is concrete: 56 Nvidia B200 GPUs, equivalent to seven eight-GPU DGX B200 systems, serving GLM-5.3 to the engineering team at a blended cost of around 0.30 dollars per million tokens. Before the switch, engineers were reportedly spending 1,000 to 10,000 dollars a month each on commercial coding tools, with strict budget limits in place to keep the bill under control. With a flat, self-hosted bill, the limits are gone.

Revolut also credits the switch with productivity gains: roughly 60 percent shorter coding time and 70 percent shorter lead time, from picking up a task to merging it, after the rollout to about 2,000 engineers. Those are Revolut's own figures, and the company has not published an independent audit of them. Keep that in mind through the rest of this article, because the caveats matter as much as the headline.

What Self-Hosted AI Coding Actually Looks Like

The model at the center of the story is GLM-5.3, from Z.ai, also known as Zhipu. It is an open-weight model, which means the weights are downloadable and anyone with enough GPUs can run them in-house without sending a single line of code to an outside API. The model is large: roughly 754 billion parameters, with about 40 billion active for each generated token, and a weight footprint near 380 GB.

Revolut only uses the open weights and runs them on its own infrastructure in the US, not through Z.ai's hosted service. That distinction matters for a regulated bank, which cannot casually send its codebase to a third-party API. The code stays inside Revolut's walls, on hardware Revolut controls.

There is a router in front of the models. GLM is text only, so it is not ideal for work involving images or interfaces. Requests with images go to models like Gemini or GPT Luna, and Claude Code and Codex remain available as the exception for the hardest work. Coding agents run in a sandbox with approval rules, with a human in the loop. The frontier models went from default to exception, not from default to deleted.

The Catches Behind the 90 Percent

The first catch is scope. The 90 percent figure covers inference costs, the per-token price of running the model. It excludes cloud spend, staffing, and monitoring, and it depends on heavy tuning. Abitbol said out-of-the-box inference was not good enough; Revolut had to tune the deployment, including FP4 precision, context-window tuning, and cache optimization, to get time-to-first-token down from over 10 seconds to 1.6 seconds. A team that downloads GLM and runs it without that work will not see Revolut's numbers.

The second catch is attribution. The productivity gains do not isolate the model change from other causes. Removing token limits alone changes behavior: engineers who were rationing AI usage start using it for everything, including Jira tickets, Confluence docs, and Slack requirements, not just code generation. Some of the 60 and 70 percent improvements may come from that freedom, not from GLM being better than Claude.

The third catch is scale. Fifty-six B200 GPUs is a serious cluster, and it only makes financial sense at Revolut's volume. A team of twenty engineers will not save money by buying seven DGX systems. The math works when you have 2,000 engineers generating enough tokens to amortize the hardware, and when your alternative is paying frontier API prices for all of them.

What Other Teams Can Actually Copy

The pattern worth copying is not the GPU cluster. It is the router. Put a cheap, good-enough model in front as the default, route the hard or visual work to frontier models, and keep a human in the loop. That architecture works at any scale, and it is the same idea behind the agent workflows spreading through the industry. The GitHub Octoverse 2026 data shows agents now start 40 percent of pull requests, and GitHub Agentic Workflows lets teams describe a job in Markdown and hand it to an agent. The router pattern fits naturally into that world.

If you want to gauge GLM's quality before committing to anything, Z.ai sells a GLM Coding Plan starting at 18 dollars a month that plugs into Claude Code, Codex, and more than 20 other coding tools. That runs on Z.ai's servers, so it tests the model, not the self-hosting. It is a cheap way to answer the only question that matters for your team: is this model good enough for our code?

The bigger story is that open-weight models have crossed a threshold. A year ago, self-hosting a coding model meant accepting a clear quality gap against Claude and GPT. Revolut's bet is that the gap has closed enough for everyday coding work, and that the savings are worth the engineering effort of closing the rest. The DeepSeek open-source coding agent points in the same direction: open tooling, running on your infrastructure, good enough for real work.

Not every team should follow Revolut. But every team that pays a large AI coding bill should now do the math. The frontier is still the best model for the hardest 10 percent of the work. For the other 90 percent, the question is no longer whether an open model can do the job. It is whether your team is big enough, and your infrastructure good enough, to make the savings real.