Revolut’s reported reduction in AI coding costs is not a claim that one model became ten times better at writing software. It is an account of changing the way a large engineering organization obtains and uses inference. The interesting part is the connection between price, utilization and the work developers decide is worth delegating.
The Register’s new report draws on an earlier conference presentation by David Abitbol, who leads AI operations at Revolut. Its figures describe that deployment, not a universal price list for self-hosted AI.
Lower token costs changed which tasks engineers attempted
The Register reports that Revolut made self-hosted GLM models the default for much of its coding work while retaining commercial frontier services for exceptions. It describes a deployment serving roughly 2,000 engineers using 56 NVIDIA B200 GPUs and a blended cost near $0.30 per million tokens. The reported spending reduction is about 90%, not a demonstrated saving available to every organization adopting the same model. Original report.
In the complete conference transcript, Abitbol argues that scarce token allowances encouraged engineers to reserve AI for difficult coding tasks. Cheaper access broadened its use to surrounding work, including searching information and dealing with issue-tracking and documentation tasks. That is his explanation of the internal experience, rather than a controlled causal study. Original conference session and transcript.
The distinction matters because a software-delivery process contains more than code generation. Reducing the time spent typing a function would not necessarily remove the delay in finding context, reviewing a change or getting it ready to deploy.
A cost comparison therefore needs to identify both what is being purchased and what work changes when the constraint is relaxed.
Hosting the weights was only the beginning
Abitbol’s presentation describes substantial serving-system tuning. His examples include choosing a practical context allowance, managing cache capacity, considering quantization and separating input processing from token generation. He also discusses routing image-related requests to other models instead of forcing a text-only coding model to handle every task. These are details of Revolut’s approach, not a prescription that another workload should copy unchanged. Conference transcript.
The useful comparison is consequently not “downloaded model versus paid API” in the abstract. It is one operating configuration against another, with actual demand and acceptable response times attached.
Hardware that is underused still has to be paid for. A configuration that fits model weights but leaves inadequate room for concurrent requests may also behave differently from a small demonstration. Those are reasons to inspect the utilization assumptions behind the claimed cost, rather than treating the headline token figure as portable.
The account is valuable precisely because it describes engineering work around the model. The claimed economics are an outcome of that system, not an intrinsic property of the model name.
Self-hosting changes responsibility, not just the invoice
The Register also reports improvements in coding and delivery times. Those are company-reported internal comparisons; they do not isolate the effects of the model choice from other workflow or deployment changes. The Register’s account.
Abitbol’s closing discussion addresses sandbox restrictions, approval rules and telemetry. Those controls are relevant because running inference on chosen infrastructure does not, by itself, determine what an agent may execute or make every output safe. His regulatory comments should not be read as a general compliance determination for other organizations.
The cost result and the governance result are therefore separate questions. One concerns the resources consumed by an operating workload; the other concerns access, oversight and the ability to investigate mistakes.
Revolut’s experience supplies a concrete case for examining self-hosted coding systems. It does not eliminate the need to account for the infrastructure, operational work and controls that made this particular deployment possible.










