How Revolut Cut Its AI Coding Bill 90% with Self-Hosted GLM
Europe’s largest private company moved about 2,000 engineers onto an open-weight model running on dedicated GPUs. Its AI bill fell to a tenth and lead time improved 70%. Here’s what Revolut did, and what the same playbook looks like for a team of 20.
Key takeaways
- Revolut serves Zhipu’s open-weight GLM-5.3 to about 2,000 engineers on 56 Nvidia B200 GPUs, at a blended $0.30 per million tokens.
- Its invoice is a tenth of what it paid Anthropic and OpenAI, and the bill is flat.
- With unlimited tokens, coding time improved 60% and lead time 70%. Budget-capped Claude Code had shown little impact.
- Faraday Machines gives a small team the same setup: GLM on hardware you control, unlimited tokens, one fixed monthly price.
The Numbers
Revolut is a European digital bank with 80 million customers in 45 markets, and an EU-regulated one. At AI Engineer NYC on October 13, 2026, David Abitbol, who leads AI operations there, explained how the company made a self-hosted open model the default for its engineers.
From Rationed Tokens to Unlimited Ones
| When | What Revolut did | Effect |
|---|---|---|
| January 2025 | Gemini, the first AI allowed to use private data | Lead time 10% faster, coding time 20% faster |
| 2025 | Claude Code at scale, with strict budget limits | Little impact |
| April 2026 | Self-hosted GLM-5.1, with unlimited tokens | Lead time 70% faster, coding time 60% faster |
| Today | GLM-5.3 on 56 B200s; Claude Code and Codex kept for exceptions | A flat bill, a tenth of the old one |
Measured on a sample of 2,000 Revolut engineers, as presented by David Abitbol.
Revolut had launched Claude Code with strict budget limits after seeing engineers spend a thousand, even ten thousand dollars a month with little return. Capped, it made little difference. What changed the numbers was taking the cap away.
Lead time improved more than coding time, and Abitbol’s explanation is the heart of this case: the bottleneck in engineering is not necessarily coding. With tokens free to use, engineers pointed AI at Jira, Confluence and their inboxes, not just at the hardest tickets. In his words, “You no longer have to save your tokens for the end of the quarter or the most difficult task.”
What It Took
- Tuning, not more GPUs. “If you just get GPUs and upload weights,” Abitbol said, “you’re going to get terrible performance.” Between April and August, Revolut cut time to first token from over 10 seconds to about 1.6 while requests per second rose from about 3 to over 7, on the same GPUs.
- Sensible limits. A 256K context window instead of a million, FP4 precision on Blackwell GPUs (about 40% less memory than FP8) and a cache hit rate above 90%.
- A response-time target. Under 3 seconds to the first token; past 5 seconds, complaints start. About 200 tokens a second is plenty, because engineers now run tasks in parallel and check back when they’re done.
- Frontier as the exception. Images go to Gemini through a router, and Claude Code and Codex stay available for the toughest work.
How a Regulated Bank Got Comfortable
- Weights, not a Chinese service. Revolut downloads GLM’s open weights and runs them on dedicated GPUs. Prompts never go to a Chinese endpoint.
- A human in the loop. Coding agents run in a sandbox with an allow-list, such as allowing grep but not gradle, which shows the regulator they aren’t autonomous agents and guards against misuse from inside.
- Data stays in approved countries, with zero data retention and detailed logs for any incident.
The Same Playbook for a Team of 20
Revolut’s lesson is simple: own the model, stop rationing tokens, keep the frontier for exceptions. Its setup is not: 56 data-centre GPUs and an AI operations team tuning them for months. Faraday Machines delivers the same idea as a product, on a Mac Studio in your office.
| Consideration | Revolut’s setup | Faraday Machines Scale |
|---|---|---|
| Model | GLM-5.3 | GLM-5.3-Flash or GLM-5.3, plus DeepSeek, MiMo and others; GLM-5.3 at full 4-bit on a custom cluster |
| Hardware | 56 Nvidia B200 GPUs | One Mac Studio M5 Ultra, 512 GB |
| Team size | About 2,000 engineers | 20–29 people, with no seat limit |
| Price | $0.30 per million tokens, plus the GPUs | $999 US a month, flat, hardware included; unlimited tokens |
| Setup and tuning | An in-house AI operations team | Installed, tuned and maintained by Faraday Machines |
| Where data goes | Revolut’s GPU providers in approved countries | Nowhere: the Mac Studio is in your office |
Smaller teams start on Starter ($299) or Growth ($599). A Mac Studio is slower per request than a B200 cluster, at about 50 tokens a second on a frontier model, which suits the run-in-parallel, check-back-later way of working Revolut describes.
Revolut’s biggest gains came from the work around the code: tickets, documents and inboxes. That is what Faraday Arc does with your company’s files, and it comes with every plan, running on the same unlimited model.
Frequently Asked Questions
What AI model does Revolut use for coding?
Revolut self-hosts Zhipu’s open-weight GLM as the default coding model for about 2,000 engineers. It started with GLM-5.1 in April 2026 and now serves GLM-5.3 on 56 Nvidia B200 GPUs. Claude Code and Codex are still available, but as exceptions for the hardest tasks.
How much did Revolut save by self-hosting?
Revolut’s AI operations lead, David Abitbol, says its blended cost is $0.30 per million tokens and its invoice is one tenth of what it was spending on Anthropic and OpenAI. The bill is also flat, so there is no end-of-month surprise.
Did engineering get faster?
Yes. On a sample of 2,000 engineers, coding time improved about 60% and lead time, from picking up a task to merging it, about 70% after Revolut began self-hosting GLM with unlimited tokens. Budget-capped Claude Code had shown little impact before that.
Can a small business do what Revolut did?
Yes, without building a GPU cluster. Faraday Machines installs open-weight models, including GLM-5.3-Flash and GLM-5.3, on a Mac Studio in your office for a fixed monthly price with unlimited tokens: $999 US a month for Scale, which suits teams of 20–29, with Faraday Arc included.
References
[1] David Abitbol, Revolut. (2026). “Shipping like a startup, regulated like a bank.” AI Engineer NYC, October 13, 2026. Transcript: ai.engineer; recording: youtube.com
[2] Faraday Machines. Plan pricing, the GLM-5.3 model profile and Faraday Arc. Available at: faradaymachines.com/pricing, faradaymachines.com/models/glm-5-3 and faradaymachines.com/arc
Stop Rationing Your Team’s AI
Tell us how your team uses AI today. We’ll show you which open model fits, which Mac Studio runs it, and what unlimited tokens at a fixed price would change.
Talk to Us