Picture 2,000 software engineers who suddenly don’t have to ration their AI anymore. No budget caps. No saving tokens for the end of the quarter. That’s what Revolut, the British fintech, says it pulled off this year.
The trick is Revolut’s self-hosted AI setup. Instead of mostly paying Anthropic and OpenAI for Claude Code and Codex, it now runs an open-weight model called GLM itself. And it says its AI coding bill fell by about 90% in under a year.
The details came from David Abitbol, who runs AI operations and infrastructure at Revolut. He spoke at the AI Engineer conference in New York last month, and The Register wrote it up on November 3. So let’s dig into what changed, and what it means for you.
So what did Revolut actually change?
Revolut didn’t jump straight to its own model. First, it rolled out Google’s Gemini across the company. That cut coding time by 20% and lead time by 10%, according to Trending Topics.
Next came Claude Code, with strict budgets per engineer. That’s where things got awkward. Abitbol said the rollout cost a lot for a modest gain, Dealroom reports. Because tokens were rationed, engineers ended up saving AI for the hard stuff only.
Then Revolut made GLM its default. GLM is an open-weight model from the Chinese AI lab Z.ai. Open-weight means the lab publishes the model’s trained “brain”, so anyone can download it and run it on their own machines. Revolut started with GLM-5.1 and later moved to GLM-5.3, The Register says.
Today, GLM is the default coding model for roughly 2,000 engineers. Claude Code and Codex haven’t disappeared, though. They’re still there as the exception for the toughest work.
Why the Revolut self-hosted AI bill dropped so much
Here’s the money part. Revolut puts its cost at around $0.30 per million tokens. A token is a small chunk of text, roughly part of a word, and it’s how AI companies charge you. Revolut says its bill is now about a tenth of what it used to pay Anthropic and OpenAI.
All of this runs on 56 Nvidia B200 chips. That’s the same as seven big servers with eight GPUs each. Since Revolut runs the model itself instead of paying per token, the bill stays flat. So there’s no nasty surprise at the end of the month.
Abitbol summed up the mood shift nicely: “You no longer have to save your tokens for the end of the quarter or the most difficult task.” In other words, engineers can just use the thing.
Out of the box, though, it was slow. Abitbol called the first inference performance “terrible”. (Inference is simply the model producing an answer.) After a lot of tuning, Revolut cut the wait for the first bit of a reply from over 10 seconds to under 2. His reason for caring was very human: “As soon as you exceed 5 seconds, my Slack inbox starts to blow up.”
Wait, coding got 60% faster?
Revolut’s own numbers say coding time fell by about 60% and lead time fell by about 70%. Lead time means the stretch from picking up a task to merging the finished code.
Those are big claims, so let’s be fair about them. They’re self-reported. Also, Revolut changed two things at once. It switched models, but it also removed the token limits. As Dealroom points out, the comparison doesn’t separate the two.
Honestly, we think the second change did a lot of the work. Abitbol said the biggest gains came from debugging and from Jira and Confluence tickets, not from writing code. “The bottleneck for engineering is not necessarily coding,” he said. That rings true for anyone who’s lost an afternoon digging through tickets.
The catch nobody puts in the headline
That 90% figure leaves some costs out. According to Dealroom, it doesn’t count cloud, staff or monitoring costs. Trending Topics also did some rough math of its own. Seven of those eight-GPU servers could run $3.2 million to $3.5 million, before power and data center bills.
Then there’s the China question. GLM comes from a Chinese lab, and Abitbol admitted there’s “a lot of stigma” around that. Revolut’s answer is to use only the open weights, on its own US infrastructure, instead of Z.ai’s hosted service. Its coding agents also run in a sandbox with approval rules. “They’re always with a human in the loop,” Abitbol said.
GLM has limits too. It’s text-only, so a router sends image tasks to Gemini instead. And let’s be plain about scale. Revolut has a big infrastructure team, and Dealroom notes smaller companies may not see the same results.
Okay, but should your team try this?
Probably not tomorrow. But the direction is clear, and it matters for teams in Armenia and other emerging markets. Per-seat AI subscriptions add up fast when your team grows.
Open-weight models give you another option. You can run them yourself, keep your data where you choose, and stop counting tokens. We wrote about this shift when Mistral released Large 4 as a free download. Revolut is the first big, real-world receipt we’ve seen for it.
So here’s a sensible middle path. Use a cheap open model for the boring, high-volume work. Then keep a frontier model on hand for the tricky stuff. That’s basically Revolut’s setup, minus the 56 GPUs.
Want the labs’ side of this? Our breakdown of Anthropic’s IPO filing shows what’s riding on customers like Revolut.
So is Revolut’s self-hosted AI the future?
For routine coding at big companies, it might be. Dealroom called it a warning for AI labs, because “good enough and cheap” is gaining ground for high-volume work. Meanwhile, AI agents now start 4 in 10 pull requests, as we covered in our GitHub Octoverse 2026 breakdown. That volume isn’t going down.
Still, Revolut’s self-hosted AI story isn’t a free lunch. It took serious hardware, a lot of tuning and a clear head about risk. Our open question: if your team could stop rationing AI tomorrow, what would you actually build with it?