# Revolut Cut AI Costs 90% and Started Shipping Faster

> Source: <https://linas.substack.com/p/revolut-ai-coding-costs-open-source-models>
> Published: 2026-11-09 01:56:17+00:00

👋 *Hey, Linas here! Every day, I break down 3 stories shaping the future of FinTech & Artificial Intelligence - plus the money movements and trends worth tracking. First time here? 400k+ FinTech and AI leaders get this daily. Join them:*

**Revolut cut its AI coding bill by 90%. Then its engineers started shipping faster.**

David Abitbol, who leads AI operations at Revolut, recently shared this at AI Engineer New York and gave **the first inside look at** **how** **Revolut runs AI coding at scale**. Revolut moved about 2,000 engineers off metered Anthropic and OpenAI APIs. They now use Z.ai’s [open-weight GLM model](https://linas.substack.com/p/glm-5-3-flash-guide), self-hosted on 56 Nvidia B200 GPUs through Together AI. The blended cost is $0.30 per million tokens, and the invoice is 1/10 of what it was. On top of that, after the switch, lead time (from picking up a feature to merge-ready code) fell by about 70% 😳

### Why Revolut’s numbers matter

*Who* is saying it matters as much as *what*. A $115 billion secondary share sale in July 2026 made Revolut [Europe’s most valuable private company](https://linas.substack.com/p/fintechpulse1087). It reported [more than $6B of revenue in 2025](https://linas.substack.com/p/fintechpulse1059), up 46%, from over 80 million customers.

Revolut is also one of the most AI-first fintechs in the world. It built [PRAGMA](https://linas.substack.com/p/revolut-pragma-foundation-model), its own foundation model for banking, with NVIDIA, and runs an in-house AI lab, [Revolut Research](https://linas.substack.com/p/revolut-research-proves-fintechs). **When a company with that much AI expertise** **changes** **how it buys AI, its suppliers, competitors, and investors should take note**.

### Cheaper AI, faster engineers

Revolut had already given its engineers the best coding agents money could buy, Claude Code and Codex. Yet, delivery speed barely moved. The 90% saving came later, and so did the productivity jump. The link between the two is how AI was priced for the people using it, and most AI companies still get that **wrong**.

**How did Revolut cut AI coding costs by 90%? 🤔** It replaced pay-per-token frontier APIs with a fixed fleet of GPUs running an open-weight model. It tuned that fleet instead of growing it, and kept frontier models for the few tasks that need them.

Below is the full breakdown of how Revolut did it, what it really costs, and what it means for anyone building, buying, or backing AI. Inside:

- **Why rationed frontier agents failed** , and the one change behind the 70% lead-time gain
- **The tuning recipe** that let the same 56 GPUs absorb 76% more demand, and the setting Abitbol calls “a scam”
- **The real cost per engineer** behind the $0.30-per-million-token headline, worked out from public GPU prices
- **How an EU-regulated bank got a Chinese-developed model through compliance** , and the architecture that made it possible
- **Where value is moving in the AI stack** , with a layer-by-layer map and due-diligence questions for investors
- **A playbook for founders** building, buying or selling AI coding tools
- **What to be skeptical about** , including where the headline numbers flatter the result
- **Bonus: the original slides from the talk** (as a downloadable PDF)

Most companies still can’t say what AI is worth to them. Our breakdown of [a16z’s State of Markets 2026](https://linas.substack.com/p/a16z-state-of-markets-2026) found that 69% of S&P 500 companies run AI in production, but only 2% track an AI metric over time. Revolut is in the 2%, with nearly two years of data across 2,000 engineers. Few case studies this year will be as useful to anyone spending or investing money in AI.