Why Revolut stopped renting AI and started running its own

According to Abitbol, Revolut tracked about 2,000 engineers over the last year on two metrics: coding time (picking up a task to writing the code) and lead time (picking up a feature to having it merged and ready to deploy).

When they rolled out commercial coding agents, cost became the worry. They saw engineers spending a thousand, even ten thousand a month, without much return, so the tools launched with strict budget limits.

So they asked a different question: how do you give engineers unlimited tokens, so they can use AI "guilt-free," for a fraction of the cost? Their answer was open source. They started self-hosting GLM, an open model they are "a big fan of" for coding, on dedicated GPUs on Together's cloud.

The results he reported after the switch: a 60% improvement in coding time and a 70% improvement in lead time. Those are the figures as spoken; one summary slide lists a 60% lead time improvement, so call it 60 to 70%.

The bigger gain on lead time is the interesting part. In his words, "the bottleneck for engineering is not necessarily coding." With unlimited tokens, engineers stopped saving AI for the hardest tasks or "end of the quarter" and used it for everything around the code: digging through inboxes, tickets and docs.

This is the slide that will get passed around. Abitbol said their blended cost is about 30 cents per million tokens, and their invoice is about a tenth of what they were spending on Anthropic and OpenAI. Billing is flat, so there is no "bill shock" at the end of the month. They still use frontier models, but as the exception, not the default.

He was also open about the work involved. Some of the figures he shared:

• About 56 B200 GPUs to serve around 2,000 engineers.
• GLM is a big model, roughly 800 billion parameters by his count.
• Aim for under 3 seconds to first token; past 5 seconds, he said, the complaints start.
• Around 200 tokens per second is enough for interactive use, and engineers who connect after 5 p.m. can see about 800 tokens per second.
• On the same 56 GPUs, tuning took time to first token from over 10 seconds to around 1.6, while requests per second went from about 3 to over 7. He summed it up as 88% better performance for 76% more demand.

The caveat he gave is real: "there's a lot of tuning that goes into running open source models." Put weights on a GPU with default settings and, he said, "you're going to get terrible performance." His advice covered context length, quantization, tensor parallelism and cache hit rates. Self-hosting is not free. It is a choice to spend effort instead of rent.

He also addressed compliance head on. As an EU-regulated bank, Revolut runs the open weights on dedicated GPUs with a US-based provider rather than calling the model maker's own hosted endpoint, keeps a human in the loop, and sandboxes what the coding tools are allowed to do.

We think this talk is a signal, not an outlier. Three things are lining up:

1. Open models are good enough for real work. A regulated bank is betting daily engineering on one.
2. Running the model yourself changes behavior. When tokens stop being metered per person, people use AI for the boring 80% of the job, which is exactly where the lead time gains came from.
3. Control is a feature. Flat costs, dedicated capacity, and your own rules about where data goes and what the AI is allowed to touch.

Not every team will run 56 GPUs. But more people and companies will want AI they choose and control, and the tools around that AI need to keep up.

This is the bet we are making with Koru, a free, open source (MPL-2.0), local-first workspace that works offline and with the AI you choose.

It doesn't lock you into one model vendor. It has a built-in MCP server, so the MCP client you already use, including one connected to a self-hosted or open model, can work with your workspace. If your company moves from a frontier API to its own GLM deployment next quarter, your workspace doesn't have to change. See the MCP docs.

The rest of Koru follows the same idea of owning your tools:

• Open source under MPL-2.0, so you can read and build on the code.
• Local-first and works offline, built in Rust with Loro CRDTs.
• Free peer-to-peer sync between your own devices, built on iroh.
• Native, not Electron, on Windows, macOS and Linux. iPhone and Android apps are in progress.
• Optional cloud sync: Koru Cloud is $9.99 a month for 100 GB, if you want an always-on copy.

Our longer-term vision is a workspace where humans and AI agents work side by side. That is where we are headed, not something shipped today. Talks like this one make us more confident that the AI in that picture will often be a model you or your company run yourselves.

The full talk is on the AI Engineer YouTube channel: How Revolut Cut AI Coding Costs 90% With Self-Hosted Open Models. All figures above are as stated by Abitbol. If you want an open source, local-first workspace that works with whichever AI you pick, download Koru.