Revolut Says Self-Hosted Chinese GLM Model Cut Its AI Coding Costs by 90%

David Abitbol, who oversees AI operations and infrastructure at the London-based neobank, described the move at the AI Engineer conference in New York, Trending Topics reported. Revolut's roughly 2,000 software engineers now use a GLM model from the Chinese lab Z.ai for coding work. Analytics India Magazine reported that the deployment runs GLM-5.3 on 56 Nvidia B200 GPUs on dedicated hardware.

According to Analytics India Magazine, Abitbol said self-hosting brings Revolut's blended token cost to 30 cents per million tokens, and that the company's bill is now one-tenth of what it previously paid Anthropic and OpenAI. Before the switch, some engineers were spending between $1,000 and $10,000 a month on tokens without a clear return, which led Revolut to impose strict usage budgets.

How Revolut Runs It

Revolut downloads only the openly published model weights and runs inference on US-based infrastructure, rather than sending requests to Z.ai's endpoints in China, both reports said. Trending Topics reported that the company's rules allow data to be processed only in the US, the EU, Britain and Israel, that providers must keep no data by default, and that coding agents work inside a sandbox with approval rules for permitted commands. "They're always with a human in the loop," Abitbol said, according to the outlet.

The company has been self-hosting open models since April 2026. Abitbol called the performance of the model straight out of the box "terrible", Trending Topics reported, and Revolut spent months tuning it. Between April and August, time to first token fell from more than 10 seconds to about 1.6 seconds, even as demand rose 76%, without adding GPUs, Analytics India Magazine reported.

Across its sample of about 2,000 engineers, Revolut measured a 60% improvement in coding time and a 70% improvement in lead time, the period from picking up a task to merging the code. Removing token limits meant engineers no longer had to ration usage on demanding tasks or near the end of a quarter.

Revolut has not dropped US models altogether. Claude Code, Codex and Google's Gemini remain available for specialised work, and because GLM handles only text, a router sends image tasks to Gemini. Abitbol described the latest Anthropic and OpenAI models as "excellent", according to Trending Topics.

The Caveats

The 90% figure covers inference costs and excludes cloud, staffing and monitoring costs, according to an analysis by StartupHub.ai cited by Trending Topics. Revolut did not say whether it owns or rents the B200 chips, and its results depend on continuous tuning and on its ability to shape demand through sandbox rules and working hours. Smaller teams trying the same approach could face high bills or slow responses.

Why It Matters

Revolut was valued at $115 billion in a secondary share sale in July, and Trending Topics reported that it has more than 80 million customers in 45 markets. A regulated bank of that size moving its default coding tool from US frontier labs to a self-hosted Chinese open-weight model is a significant data point for enterprise buyers weighing the cost of AI coding agents, and for Anthropic and OpenAI, which count coding among their fastest-growing revenue lines.

The case also shows the compliance pattern some firms are adopting for Chinese open models: using the weights while keeping inference and data inside approved jurisdictions. Revolut chief executive Nik Storonsky said in October that the company is also developing its own models trained on customer transaction data, Analytics India Magazine reported.