BUSINESSCompanies · · 2 min read · Source: The Register
Revolut: self-hosted models cut AI coding costs by about 90%
Digital bank Revolut says it cut the cost of its AI coding assistants by roughly 90% in under a year. Instead of relying mainly on commercial services such as Claude Code, it now runs open models on its own hardware, The Register reported on November 3, 2026.
My tool tip · AdWispr Flow – AI dictationTry it free
Key facts
- Revolut says it cut its AI coding costs by around 90% in under a year.
- The open GLM model runs on 56 Nvidia B200 GPUs and costs about $0.30 per million tokens, according to Revolut.
- Around 2,000 Revolut engineers use the AI tools; The Register reported the figures on November 3, 2026.
What Revolut did
The figures come from a talk by David Abitbol, who leads operations and infrastructure at Revolut, at the AI Engineer conference in New York. Revolut rolled out AI coding tools to around 2,000 engineers: Gemini in early 2025, Claude Code in early 2026. Because some engineers were spending $1,000 to $10,000 a month without clear returns, Abitbol said, the company changed course.
The default is now an open model (first GLM 5.1, now GLM 5.3) running on 56 Nvidia B200 GPUs. According to the talk, that costs around $0.30 per million tokens – roughly a tenth of commercial coding services. A router sends image requests to other models.
The reported results
- around 60% less coding time and 70% shorter lead time (Revolut's figures)
- first token in under 3 seconds; complaints started once latency passed 5 seconds
- the biggest gains came from debugging and tickets, not from writing code
Note: these are claims from the talk and haven't been independently verified.
Why this isn't for everyone
56 high-end GPUs and a team to run and tune them only pay off at scale. For small teams, subscription services usually remain simpler. The story is still useful because it shows how fast company AI usage can get expensive without budget limits.
What this means for you
- Without limits, AI costs per engineer can quickly reach four figures a month.
- Self-hosted open models have become a real alternative to subscriptions for large companies.
- Revolut says the biggest gains were in debugging and tickets, not in writing code.
How to use this
Pull last month's usage and cost report per person from your AI provider. Anyone far above average shows you where a limit or a conversation about value makes sense.
Luigi's take
The 90% sounds spectacular, but it applies to a company with 2,000 engineers and its own infrastructure. For most businesses the more useful lesson is different: Revolut first measured what AI cost and delivered, then decided. Many companies hand out AI licenses without ever looking at the bill per person.
My advice: set a per-person budget for every AI tool on your team and check after three months what it's actually used for. The value often lies somewhere unexpected – at Revolut, in tickets and debugging. Only then is it worth asking about a cheaper model.
Source
More sources
- AI Engineer NYC 2026 – Vortrag von Revolut, auf dem die Angaben beruhen
This article was researched and written with AI assistance. Responsible: Luigi Lugmayr.
