Revolut says it cut its AI coding bill tenfold by self-hosting a Chinese model

Some Revolut engineers were spending between $1,000 and $10,000 a month on coding assistants, for a benefit judged to be low, The Register reported on November 3. The British neobank has therefore changed approach: it now hosts its AI itself.

David Abitbol, Revolut's head of operations and infrastructure, detailed the shift at the AI Engineer conference in New York. Revolut used Google's Gemini from early 2025, then Anthropic's Claude Code from early 2026, with strictly capped budgets. The default tool is now GLM, an open-weight model from the Chinese company Z.ai, which the bank hosts itself. The bill would be roughly ten times lower than with Anthropic's or OpenAI's services.

56 NVIDIA chips for 2,000 engineers

Revolut started with GLM-5.1 and now serves GLM-5.3 to about 2,000 engineers, on 56 NVIDIA B200 graphics processors. GLM-5.1 is massive: about 754 billion parameters, of which 40 billion are used for each token generated, according to The Register. The average cost is around $0.30 per million tokens, the fragments of text the model reads and writes.

Abitbol stresses that tuning, more than extra hardware, made the operation viable. Revolut compressed the model's weights to 4 bits, separated the processing of the question from the generation of the answer, and adjusted the context length. On the same chips, the time before the first word appears dropped from 14.26 seconds to 1.68 seconds. The goal is to stay under 3 seconds with a context of about 256,000 tokens, at nearly 200 tokens per second. A one-million-token window is useless in practice, he says.

The gain doesn't come from where expected

According to him, the main benefit doesn't come from writing code. It comes from debugging and from handling Jira and Confluence tickets, which are handed to the model now that no one rations requests. Revolut reports a 60% improvement in coding time and a 70% improvement in delivery time. However, the conference page specifies that this comparison does not separate the effect of the model from other changes.

Commercial models haven't disappeared, though. GLM only handles text: an automatic router sends tasks that include images to Gemini or GPT. Code generation, for its part, stays on GLM.

What the figures don't say

These results are Revolut's own, presented on stage and not independently verified. Above all, Revolut didn't detail what its $0.30-per-million-token cost covers: it's unknown whether it includes the 56 chips, the energy and the team that runs them. The net savings therefore remain unknown.

Then there's the model's origin. Abitbol acknowledges the privacy and security concerns raised by Chinese models. For him, the main risk is querying these models through services hosted outside one's control, in China; running them in-house or in a regional cloud is far less worrying. Open models specialized in code are multiplying, like the one DeepSeek released in October. Revolut's example shows they can be used at scale, provided you have the infrastructure skills to run them.