Revolut moves 2000 engineers to self-hosted GLM and cuts AI coding costs by 90%

AI coding bills are growing faster than teams can work out what value they are getting. Revolut, the British fintech with tens of millions of customers, took a radical approach: it cut spending on coding assistants by approximately 90% in less than a year. That is almost a tenfold difference, making this a particularly interesting case for a financial company with strict data requirements.

What Revolut did

The bank dropped Claude Code and Codex and switched to the open GLM model from Chinese lab Z.ai. The model runs on Revolut's own GPUs, meaning engineers' code and prompts do not go to an external provider. Judging by the source article's headline, the cost is around $0.30 per million tokens, with approximately 2000 engineers moving to the new setup.

David Abitbol, Revolut's head of operations and infrastructure, presented the figures at the AI Engineer conference in New York. An important caveat: these are the company's own estimates, and no one has independently verified them. The 90% figure should therefore be treated as a claimed result, not an established fact.

Where the savings come from

The material available to us does not include details of the calculation, so what follows is our interpretation, not a recap of the presentation.

Paid assistants charge by the token or by subscription seat. Agent workflows consume enormous quantities of tokens: the model reads a repository, edits files, runs tests, and reads the output again. With thousands of engineers, the bill rises with usage, and the better people become at using the tool, the harder it hits the budget.

Running your own hardware turns a variable expense into a largely fixed one. The GPUs have already been purchased or rented, and each additional request incurs electricity and depreciation costs. A simple example: if an agent uses several million tokens on a task, at $0.30 per million that comes to roughly a dollar. The same task using a premium model through a commercial API could cost several times more.

But the savings only materialize when GPU utilization is high. An idle cluster quickly eats up all the savings.

What remains unclear

Several gaps in the story need to be considered before copying the approach:

• Quality. We do not know how GLM compares with Claude and Codex on Revolut's actual tasks, or how often engineers have to rework its output.
• Total cost. It is unclear whether the $0.30 per million tokens includes GPU purchases, electricity, and the work of the team maintaining inference infrastructure.
• Reasons for the choice. For a bank, self-hosting is likely to involve data security and compliance as well as price. This is our assumption.
• Replicability. Revolut has a strong infrastructure team. A startup with 30 people may not have the same capacity.

The practical takeaway

The main lesson is not that open models have beaten closed ones. It is that AI coding costs can be measured and optimized like any other infrastructure expense.

What teams should do:

• Measure cost per accepted pull request or completed task, rather than per token. A cheap model that requires frequent rework may end up costing more.
• Run a pilot with 20–50 engineers and compare an open model with your current tool on your own tasks, rather than public benchmarks.
• Estimate future GPU utilization before buying hardware. Self-hosted inference pays off under a steady workload.
• Consider a hybrid approach: route routine work to your own model and leave complex architectural tasks to a leading closed model.

If spending on assistants is already a noticeable budget item, Revolut's case deserves at least a place in your scenario spreadsheet. Test it against your own data rather than taking the claimed 90% at face value.