What Revolut's self-hosted GLM shows about open-weight economics

4 Nov 2026 · Private Catalyst has no relationship with Revolut. Figures are Revolut's own, from a public talk.

What Revolut said

David Abitbol leads AI operations at Revolut, which he describes as a digital bank in Europe with 80 million customers. At AI Engineer NYC 2026 he said: "to serve 2,000 engineers, we roughly need 56 B200. And that's to host GLM 5.3" (transcript, video). On cost: "blended cost of token is 30 cents per million. And so our invoice is one-tenth of what we were spending on Anthropic and OpenAI." We haven't seen those numbers audited. Read them as one company's report of its own bill.

When tokens stop being rationed

The price matters less than the shape of the bill. Abitbol says Revolut first launched hosted coding tools with strict budget limits, then asked how to give engineers tokens to use "guilt-free" at a fraction of the cost. With fixed GPUs, "the billing is flat, we don't have any bill shock any of the month", and engineers "no longer have to save your tokens for the end of the quarter or the most difficult task".

That changes what people use the model for. Abitbol's point is that the bottleneck for engineering "is not necessarily coding": time goes to inboxes, Jira and Confluence. Metered tokens get saved for hard problems. Unmetered ones get spent on the dull work that actually slows a team down.

Read the productivity numbers carefully

The talk shows a chart of about 2,000 engineers' lead time and coding time over a year. Abitbol says that around April 2026, when Revolut started self-hosting GLM-5.1, "we see 60% improvement in coding time, 70% improvement in lead time". That is Revolut's internal metric for the period after self-hosting began, as Abitbol presents it. It isn't a controlled comparison of GLM against hosted models, and he draws a narrower lesson from it: lead time moved more than coding time.

Self-hosting is engineering work

Abitbol is blunt that you can't just load weights onto GPUs: "you're going to get terrible performance". He says that in April time to first token was over 10 seconds at around 3 requests per second; by August, on the same number of GPUs, it was around 1.6 seconds at over 7 requests per second. The cost advantage depends on that tuning, and on someone owning it.

Frontier models stay, as the exception

Revolut didn't drop hosted models. "Obviously we keep using codex, code code. The latest model are fantastic... But we keep them as exception." GLM is text only, so Abitbol suggests a local router that sends image recognition to another model and keeps GLM for the main task. The pattern is a default and an exception, not all or nothing.

What a regulated firm can take from it

Abitbol says Revolut, as an EU-regulated bank, falls under a strict compliance regime. His answer for coding agents was a sandbox with rules for which operations are allowed, so a human stays in the loop, and running downloaded weights on dedicated GPUs rather than calling an endpoint abroad. Those are controls you own, not promises you rent.

The same reasoning is why we install models on hardware our clients control: a fixed cost you can plan for, prompts that stay on your network, and reasoning you can log and review. A bank with 2,000 engineers and a firm with 50 staff will size things very differently, so treat Revolut's numbers as a direction, not a quote.

Want to know what your own token bill would look like on fixed hardware? We'll map the workloads and size the machine before you buy anything. Start here →