Revolut engineers no longer have to ration AI tokens, thanks to a self-hosted open-weight model. The productivity jump that followed is bad news for every AI lab pricing its way toward an IPO.
Ask Revolut’s engineers to name their go-to coding tool and Claude Code and Codex probably won’t come up first anymore. David Abitbol, who leads AI operations at Revolut, explained why on a recent AI Engineer podcast appearance. Roughly 2,000 Revolut engineers now write code with a self-hosted GLM model running on 56 Nvidia B200 GPUs, at a blended cost of about 30 cents per million tokens. That’s close to a tenth of what Revolut was paying Anthropic and OpenAI. The two paid tools have been pushed to the exception list: the ones engineers reach for only when the self-hosted setup genuinely can’t do the job.
The cost cut is the headline number, but Abitbol’s own framing of why it worked is the part worth sitting with. Engineering lead time improved 70% after the switch, according to Abitbol, and that gain showed up specifically once tokens stopped being metered. Think about the incentives. When every API call carries a per-token bill from Anthropic or OpenAI, engineers ration it without thinking twice. Save the model for the gnarly ticket, skip it for the small stuff, treat it like a scarce resource because it is one. Drop the marginal cost to near zero and that instinct disappears. Engineers stopped reserving AI for hard problems and started pointing it at everything: pull requests, Jira tickets, Confluence pages, their own inboxes, the surrounding busywork that never touched a model before.
Lead time, from opening a ticket to shipping the fix, improved more than raw coding speed did, which suggests the gains came from automating the work around the code.
Getting to that price wasn’t just a matter of buying GPUs and pointing a model at them. According to Abitbol, Revolut’s out-of-the-box inference performance on the self-hosted setup was bad. So the team spent real engineering effort tuning context limits, quantisation and parallelism to squeeze more throughput out of the same 56 B200s. The payoff was an 88% performance improvement that let the system absorb 76% more demand without adding a single GPU. Time to first token, how long an engineer waits before the model starts responding, fell from more than 10 seconds to 1.6 seconds in the process. It’s no weekend project. It’s the kind of infrastructure investment only a company with Revolut’s scale and engineering headcount can justify. That’s exactly why it matters that a bank, not a research lab, is the one proving it out in production.
Revolut isn’t a hobbyist shop running this as a side experiment. The company has around 2,000 engineers on the self-hosted path, and this system runs as their actual daily coding tool, not a pilot programme quietly compared against a control group. Revolut has also been talking up Nvidia hardware more broadly this year, including work on a foundation model called PRAGMA built on Blackwell Ultra GPUs that it discussed at Nvidia’s GTC Berlin conference in October. The GLM coding setup is a separate, narrower bet. But it fits a pattern: a fintech with the scale to run its own inference stack is choosing to own more of it. It isn’t renting intelligence by the token.
Why this is a pricing problem, not just a cost story
Frankly, this should worry anyone banking on frontier-model pricing holding up through an IPO roadshow. Anthropic and OpenAI have built their revenue models on the assumption that quality commands a premium large enough to keep enterprise customers paying metered rates indefinitely. Revolut just showed a different path. An open-weight model, GLM, run on owned hardware, delivers the productivity win, and not because it’s smarter than Claude or GPT: it’s cheap enough that nobody has to think about the bill before using it. If engineering lead time improves 70% simply because the token meter stops running, the quality gap that frontier labs charge for starts looking like the wrong thing to be charging for.
Claude Code and Codex remain at Revolut for jobs the self-hosted model can’t handle cleanly, and frontier models may become the add-on for the hardest 10% of problems.
Also read: Mistral Is Racing OpenAI and xAI on a Fraction of Their GPUs • Looking for an Eventbrite Alternative? How Small Organizers Are Selling Tickets Without Juggling Five Tools • Nvidia tells cloud clients to cap China business at 49% of their services
This article is posted in AI News, check it out for more related stories.
Almost there. Sign in and your reply posts straight away.
