Revolut Moves 2000 Engineers to Its Own GLM Deployment: Tokens Are 10x Cheaper

Almost every team that has rolled out AI coding assistants ends up having the same conversation sooner or later: the token bill is growing faster than anyone would like, and someone says, "Let's only use it for hard tasks." Revolut, if a recent post on X is accurate, found another way out. It installed its own GPUs, ran the open GLM model on them, and made it the default tool for about 2000 engineers.

What exactly Revolut did

According to a retelling of the post, the setup looks like this. The company runs 56 Nvidia B200 GPUs in its own infrastructure. GLM, an open model that can be self-hosted, runs on them. The price comes to roughly $0.30 per million tokens, 10x lower than what the company paid Anthropic and OpenAI. Claude Code and Codex were not dropped: they moved to an exceptions list for tasks that need the strongest model, and are no longer the default.

The author of the post cites one more figure: development lead time, meaning the time from starting work on a task to shipping it, improved by 70%. The retelling does not say how it was measured, so treat this number as a claim, not a verified benchmark.

Why economics won, not quality

The most interesting part of the story is the explanation for the effect. By the author's assessment, GLM is slightly weaker than the top closed models. But AI used to be rationed: it was hooked up to hard tickets, while routine work was done by hand to keep the bill from ballooning. Once tokens became nearly free for the engineers themselves, they began running the assistant on everything: small fixes, tests, log analysis, code review, documentation.

Think of a taxi versus your own car. The taxi is better and more comfortable, but you only take it when it's "worth it." Your own car is a bit more basic, but you drive it everywhere without a calculator in your head. In total, you cover more ground and get more done. A ration-coupon model loses to a no-limits model, even if the former is smarter.

Who this suits and who it doesn't

Some sobriety is needed here. First, this is a retelling of a post, not an official Revolut statement. It gives no details on the GLM version, on how the company arrived at $0.30 per million tokens, or on whether that figure includes the purchase of the B200s, electricity, and the work of the engineers who maintain it all. Those cost items are what most often break the neat "our own cluster is cheaper than an API" math.

Second, the setup works at a large company with a strong infrastructure team. The author of the post says so himself: for small teams without their own GPUs and specialists, an API is still simpler and cheaper. For an individual developer, buying a Mac Studio, for example, to run local models would pay for itself compared with paying for an API in about 10 years.

The story is also unwelcome for labs that build their business on a high token price: if large customers start keeping open models as the default, closed flagships will be left with the role of an expensive "specialist for hard cases."

What to do about it

If you have hundreds of engineers and your own DevOps, it makes sense to estimate three things: how many tokens actually go to external providers today, what share of tasks can tolerate a slightly weaker model, and how much it costs to maintain a cluster, including the people. Then set up routing: an open model by default, with hard cases sent to Claude Code or Codex, either manually or automatically.

If you are a small team, the main lesson from Revolut applies even without B200s: check whether you are rationing tokens in places where AI could take over routine work. Sometimes a cheap model with no restrictions is more useful than an expensive one you are allowed to call once a day.