# How Revolut Cut AI Coding Costs 90% With Self-Hosted Open Models

David Abitbol, Revolut | AI Engineer NYC 2026 | 11:47

Source: https://www.youtube.com/watch?v=Ok6QdZHCcI8
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/how-revolut-cut-ai-coding-costs-90-with-self-hosted-open-models
Published: 2026-10-29
Tags: gpus, inference, open-models, quantization

## TL;DR
- Revolut found that self-hosted open models improved coding time by about 60% and lead time by about 70% across 2,000 engineers.
- Running GLM on dedicated GPUs brought the blended token cost to 30 cents per million, about one-tenth of Revolut's previous Anthropic and OpenAI spending.
- Tuning the same 56 GPUs reduced time to first token from more than 10 seconds to about 1.6 seconds while demand increased by 76%.

## Summary
David Abitbol describes how Revolut evaluated AI coding tools across 2,000 engineers and moved much of its workload to self-hosted open models. Gemini initially reduced coding time by 20% and lead time by 10%, while Claude Code had little early impact and raised cost concerns. After Revolut began self-hosting GLM, coding time improved by about 60% and lead time by about 70%. Abitbol says the larger gain in lead time came from reducing work around coding, such as searching inboxes and managing Jira and Confluence. He explains the infrastructure, latency targets, GPU tuning, quantization, tensor parallelism, cache hit rates, routing for vision tasks, and OpenAI-compatible APIs. The talk also covers compliance for an EU-regulated bank. Revolut uses US-based infrastructure with downloaded model weights, human approval through sandboxes, zero data retention, and detailed telemetry.

## Key ideas
### Self-hosted GLM produced the largest improvement in Revolut's engineering measures
[01:45](https://www.youtube.com/watch?v=Ok6QdZHCcI8&t=105s)
Revolut first introduced Gemini in January 2025 and later experimented with Claude Code. Gemini reduced lead time by 10% and coding time by 20%, while Claude Code initially had little impact. When Revolut began self-hosting open models around April, the company saw about 60% improvement in coding time and about 70% improvement in lead time. Abitbol says the company had imposed strict budget limits after seeing engineers spend from $1,000 to $10,000 per month on AI with limited return.

### Lead time improved more because AI also removed work around coding
[03:20](https://www.youtube.com/watch?v=Ok6QdZHCcI8&t=200s)
The larger gain was in lead time, which measures the period from picking up a feature to merging it for deployment. Abitbol says coding itself was not the only bottleneck. Engineers also spent time digging through inboxes and working with Jira and Confluence. Unlimited tokens let them use AI for those tasks without saving their allowance for difficult work or the end of a quarter.

### Serving 2,000 engineers requires latency targets more than extreme throughput
[04:24](https://www.youtube.com/watch?v=Ok6QdZHCcI8&t=264s)
Revolut uses roughly 56 B200 GPUs to host its large GLM model for 2,000 engineers. Abitbol says the target is less than three seconds to the first token. Once latency exceeds five seconds, his Slack inbox starts to fill with complaints. Interactive throughput is about 200 tokens per second, which he says is sufficient because engineers increasingly start tasks in parallel and check them when they finish. At quieter times, engineers can reach about 800 tokens per second.

### Tuning the serving stack improved the same hardware by 88%
[05:46](https://www.youtube.com/watch?v=Ok6QdZHCcI8&t=346s)
Revolut kept the GPU count at 56 while improving request handling. At the beginning, demand was around three requests per second and time to first token was above 10 seconds. By August, demand was above seven requests per second and time to first token was about 1.6 seconds. Abitbol reports an 88% performance improvement while demand rose by 76%. He warns that simply loading model weights onto GPUs produces poor performance without workload-specific tuning.

### Context length, quantization, parallelism, and cache hits affect cost and speed
[07:05](https://www.youtube.com/watch?v=Ok6QdZHCcI8&t=425s)
Abitbol argues that very long context windows are usually unnecessary for coding workloads. He says 256K tokens is generally enough even when a model can serve one million tokens. For Blackwell-generation GPUs, he recommends FP4 rather than higher precision, and says moving from FP4 to FP8 saves about 40% of GPU footprint without precision loss. Tensor parallelism also affects KV cache capacity, so TP8 may work better than TP4 even when the model fits on TP4. Coding workloads should have more than 90% cache hit rate.

### A text-only model can be paired with a router for vision tasks
[08:00](https://www.youtube.com/watch?v=Ok6QdZHCcI8&t=480s)
GLM is text-only, which limits some uses. Revolut chooses the stronger coding model and routes image recognition elsewhere when needed. Abitbol describes a local router that can send image tasks to Gemini or another vision model while retaining the main model for the rest of the task. Revolut's UI designer uses a vision model for interface work. The endpoints expose an OpenAI-compatible API, so they can be used with coding tools that support that interface.

### Human approval and data controls support regulated use
[09:17](https://www.youtube.com/watch?v=Ok6QdZHCcI8&t=557s)
Abitbol says Revolut's coding tasks are not treated as autonomous agents because a human remains in the loop. A sandbox controls which operations the system may perform, with approval rules for disallowed actions. The sandbox also limits misuse by employees and helps address prompt injection. Revolut uses model weights on US-based infrastructure rather than sending inference to China. Abitbol also mentions approved data locations, zero data retention from providers, and detailed telemetry for investigating incidents.

## Notable quotes
- "We see 60% improvement in coding time, 70% improvement in lead time." (03:20)
- "The bottleneck for engineering is not necessarily coding." (03:20)
- "As soon as you exceed 5 second, my Slack inbox starts to blow up and so we know that's the limit." (04:29)
- "We've had 88% improvement in performance for 76% increased demand." (06:44)
- "You start by creating a sandbox that guarantees that." (10:09)

## Tools & references mentioned
- Revolut
- Gemini
- Claude Code
- Claude Code Sonnet 3.7
- GLM
- Anthropic
- OpenAI
- Codex
- OpenAI-compatible API
- Jira
- Confluence

## Who should watch
- You are deciding whether self-hosting an open coding model can reduce API spend at a large engineering organisation.
- Your team needs practical serving targets for GPU count, time to first token, quantization, tensor parallelism, and cache hit rate.
- You work in a regulated company and need a concrete approach to human approval, sandboxing, data location, retention, and telemetry.

## Related talks

- [When Running Your Own AI Models Beats Paying for APIs](https://aietalks.com/talks/when-running-your-own-ai-models-beats-paying-for-apis) (Steven Latré, imec, 18:30)
- [The AI Bill Is Too Damn High: Tokenomics 101](https://aietalks.com/talks/the-ai-bill-is-too-damn-high-tokenomics-101) (Eugene Cheah, Featherless AI, 55:25)
- [What Makes Open Models Fast in Production](https://aietalks.com/talks/what-makes-open-models-fast-in-production) (Sujee Maniyam & Dylan Bristot, Nebius, 20:32)
- [The Desktop Frontier](https://aietalks.com/talks/the-desktop-frontier) (Ahmad Osman, Osmantic, 18:02)
- [Trends Across the AI Frontier](https://aietalks.com/talks/trends-across-the-ai-frontier) (George Cameron, Artificial Analysis, 17:52)
