Revolut went headless. The step after is owning the model

Revolut has just made the clearest case yet for open-source AI in a large company. In under a year it moved the default AI coding model for about 2,000 engineers from Claude Code and Codex to GLM, an open-weight model it serves on dedicated GPUs, and reports that its bill fell by about 90%. It is a turning point worth studying closely, because it shows both how far a company can get by owning its inference and where the bigger advantage still lies.

The problem every company now has

Most companies rent the same few frontier models. That has two consequences that get worse with scale. The first is cost: every prompt is metered, so the more people use AI, the faster the bill grows. Before the switch, some Revolut engineers were spending between $1,000 and $10,000 a month each on AI coding tools, and the company had to put strict budgets in place (The Register). Budgets controlled the bill, but they also stopped people from using the tools freely.

The second consequence is quieter. When every company sends its work to the same model, nobody's AI knows anything special. Your competitors get the same answers you do. The only thing that can make a model yours is what it learns from your business, and that is exactly what renting does not give you.

We have spent the last four years implementing AI inside established companies, and we see the same pattern everywhere: AI spend rising, usage rationed, and the company's own knowledge left out of the model entirely.

What Revolut actually did

Revolut's operations and infrastructure lead, David Abitbol, described the move at the AI Engineer conference in New York (talk video). The main pieces:

• Open model as the defaultGLM-5.1 from April 2026, later GLM-5.3, from the developer Z.ai. Frontier models stay available for the tasks where they are clearly better.
• Its own serving56 Nvidia B200 GPUs, with the weights downloaded and run on dedicated infrastructure, so no request goes to the model maker.
• A router in fronttools talk to one OpenAI-compatible endpoint, and image tasks are routed to vision-capable models. That is what "headless" means here: the tools people use are separated from the model behind them.
• Serving tuned before hardware addedabout 80% of the performance gain came from configuration (quantisation, caching, context limits), not from more GPUs.

These are Revolut's own figures. The direction is what matters: once the bill stopped growing with every prompt, engineers used AI for debugging, tickets and documentation as well as code, and that is where Abitbol saw the biggest gains.

What Revolut has not done yet

Revolut now owns its inference. It does not yet own its intelligence. Nothing in the talk or the coverage describes training the model on Revolut's own code, tickets or decisions. The model knows what every copy of GLM-5.3 knows, plus whatever is pasted into the prompt at the moment of the request.

That last step is harder than serving a model, and the hard part is not the training. It is the data. A company's knowledge lives in mail, chat, documents, tickets and system records, written for people, duplicated, contradictory and scattered across a dozen SaaS tools. Before any of it can teach a model, it has to be found, cleaned, labelled and structured, and that work has to keep going as the company runs.

Where OpenLazarus fits

OpenLazarus works like a librarian for the company. The applications teams use every day run on one operating system, so the decisions, interactions, processes and outcomes they produce are captured and structured as work happens. That turns the company's operations into data a model can learn from, prepared for training on open-source architectures.

From there, the path is the one Revolut started, taken one step further:

• Routeevery request goes through our AI Router, to a frontier model, an open model or a model trained for you (how headless AI is priced).
• Own the datayour data lives in one place that you own, not inside your software vendors.
• Trainwe fine-tune an open-weight model on your prepared data, so its answers reflect how your company actually works.
• Keep swappingbecause the tools are headless, a better open model next year is a configuration change, not a rebuild.

OpenLazarus is a European company, with European servers, which matters to the many companies that cannot send their knowledge to a model provider outside the EU.

The path, in five steps

• Measurerecord AI cost, tokens and latency per team for a month, so you know what you are actually paying for.
• Routeput one endpoint in front of every model, so switching becomes a setting.
• Move the defaultrun the bulk of the work on an open model and keep frontier models for the hardest tasks.
• Organise your knowledgebring your data into one structured, owned layer and build an evaluation set from real work.
• Train and routefine-tune an open model on that data, send it the work it does best, and keep measuring.

Revolut has shown that the first three steps pay. The last two are where a company stops renting the same intelligence as everyone else and starts owning its own.