# Revolut Says Self-Hosting GLM Cut Its AI Coding Costs by About 90%, and Claude Is Now the Exception

At the AI Engineer conference, the fintech's infrastructure lead described moving about 2,000 engineers onto an open-weight model on 56 Nvidia B200 GPUs. The details that matter are in how it serves the model, what it still sends to other models, and what the 10x figure compares.

By Priya Nair (AI Industry Reporter) — published 2026-11-09, updated 2026-11-09
Source: https://aiscoutdaily.com/news/revolut-self-hosted-glm-cuts-ai-coding-costs-90-percent

## TL;DR
Revolut says it cut AI coding inference costs by roughly 90% in under a year by making self-hosted Z.ai GLM models its default instead of commercial services like Claude. About 2,000 engineers use the setup, served on 56 Nvidia B200 GPUs at about $0.30 per million tokens, and the company reports a 60% improvement in coding time and 70% in lead time. GLM is text-only, so vision requests go to Gemini or GPT Luna, and Claude Code and Codex remain as exceptions. The figures are the company's own, from a conference talk.

Revolut's David Abitbol, its ops and infrastructure lead, told the AI Engineer conference in New York last month that the company had made self-hosted models its default for coding, as The Register reported on 3 November. The headline number is a roughly 90% cut in inference costs in less than a year. The useful part of the talk is the engineering that sits behind it.

## The setup in numbers

| Item | What Revolut reported |
| --- | --- |
| Users | About 2,000 engineers |
| Model | Z.ai GLM, starting on GLM-5.1 and moving to GLM-5.3 |
| Model size (GLM 5.1) | About 754 billion parameters, 40 billion active per token, roughly 380 GB of GPU memory for the weights |
| Hardware | 56 Nvidia B200 GPUs, the equivalent of seven eight-GPU DGX B200 systems |
| Cost | About $0.30 per million input and output tokens, around a tenth of Codex or Claude Code |
| Latency target | Under 3 seconds to first token at about 256,000 tokens of context; about 200 tokens a second |
| Productivity | 60% better coding time, 70% better lead time |

*From The Register's report of Abitbol's talk. All figures are Revolut's own and haven't been independently checked.*

## How the cost came down

- Budget limits first. Before self-hosting, Revolut had Gemini from early 2025 and Claude Code from early 2026. Abitbol said engineers were spending $1,000 to $10,000 a month with little return, so it launched with strict token budgets.
- Serving tuned, not just scaled. Abitbol said the biggest gains came from tuning hyperparameters, NVFP4 quantization, splitting prefill from decode, and tuning context length, not from adding GPUs.
- Context kept short. He said a one-million-token window can be served but 256,000 is what teams need: "Long context is a scam."
- Flat billing. "The billing is flat, we don't have any bill shock at the end of the month, and the frontier is the exception."

## What it still doesn't replace

GLM is text-only, so a model router sends vision queries to models like Gemini or GPT Luna while code generation stays on GLM. Claude Code and Codex remain available as exceptions. Abitbol also said the largest gains came from letting the model handle debugging and Jira and Confluence tickets, which engineers stopped rationing once tokens were cheap.

## What the 10x figure compares

The Register's headline says 10x and its text says roughly 90%, which are the same claim. Both compare Revolut's earlier spend on commercial services with its self-hosted cost per token. Abitbol's own account, that cheap tokens made engineers use far more of them, means total usage differs between the two periods, so the saving is per token and not a like-for-like bill. The report doesn't give the hardware or staffing cost of running 56 B200s.

## How the model compares

Thinking Machines' model card, published on 5 November, lists third-party index scores for several models. On the Artificial Analysis Index (v5.0) it shows GLM 5.3 at 31.7 and Claude Haiku 5.5 at 36.5. Those are the lab's own tables for a different product, so treat them as context and not a test of Revolut's setup.

## What this means if you pay for AI coding tools

- Per-seat pricing hides usage. Revolut's engineers were spending thousands of dollars a month before limits, so check what your heaviest users actually cost.
- Self-hosting only pays off at scale. A fleet of 56 B200s is within reach of a fintech with 2,000 engineers, not a five-person team.
- A router is the realistic design. Cheap open models for routine code, a frontier model for hard or visual tasks.
- Check where the model runs. Abitbol said the data concern is with Chinese-hosted APIs, and that running the weights on-premises or on regional cloud avoids it.

## Sources
- [The Register: British fintech Revolut slashes AI coding costs 10x by cutting out Claude and going self-hosted (3 Nov 2026)](https://www.theregister.com/ai-and-ml/2026/11/03/british-fintech-revolut-slashes-ai-coding-costs-10x-by-cutting-out-claude-and-going-self-hosted/5307345)
- [Thinking Machines Lab: Inkling 2 Small model card (third-party index scores for GLM 5.3 and Claude Haiku 5.5)](https://thinkingmachines.ai/model-card/inkling-2-small/)