- Revolut cut Claude for self-hosted Z.ai models and slashed inference costs by 90%
- 70% improvement in lead time productivity, 60% up for coding time productivity
- The model is reportedly running on 56 Nvidia B200 GPUs with 380GB+ memory
Revolut has become one of a growing list of companies looking to reduce its reliance on Claude, and in doing so, it’s reportedly cut inference costs by as much as 90% in less than a year.
But the London-headquartered digital bank hasn’t cut AI out altogether, instead it’s replaced those costly workloads with self-hosted, open-weight models from China’s Z.ai.
Revolut cuts Claude for self-hosted, open-weight Chinese models
Revolut seems pretty au-fait with the world of AI, having previously experimented with Google‘s Gemini before opting to use Anthropic’s Claude, but some heavy users were said to be spending thousands of dollars per month on tokens without generating the equivalent ROI.
“We were highly skeptical about the cost of AI,” Revolut Infrastructure Lead David Abitbol said in a conference last month . “When we launched it we launched it with strict budget limits.”
Speaking separately at GTC Berlin, Abitol also noted that Revolut’s self-hosted option resulted in a 70% improvement in lead time productivity, and a 60% improvement in coding time productivity. He also praised Z.ai’s GLM models for their data privacy and auditability.
Are self-hosted models cheaper to run?
It seems that GLM-5.1 brought inference costs down to around $0.30 per million input/output tokens, or around one-tenth the cost of similar commercial costs. Now, Revolut can operate its own inference infrastructure to manage its own capacity.
However, with great flexibility comes great cost. While costs associated with direct model usage have been slashed, self-hosting requires significant outlay in terms of hardware and infrastructure. GPUs, maintenance, electricity and maintenance and some of the key areas that will likely have seen a major increase in expenditure.
The bank now reportedly has 56 Nvidia B200 GPUs, plus the model alone needs around 380GB of GPU memory and that excludes the estimated 6 to 26GB of memory per user for context retention.
Ultimately, it’s proof that <a href="https://myappsplus.com/3-artificial-intelligence-ai-stocks-that-are-better-and-safer-buys-than-anthropic/” title=”3 Artificial Intelligence (AI) Stocks That Are Better and Safer Buys Than Anthropic”>artificial intelligence is a costly beast and deploying it aimlessly in a hope to save time is a poor strategy. But with more companies ditching commercial AI from the likes of Claude in favor of self-hosting, it begs the question whether using less advanced models for mundane work and even self-hosting over renting could improve the economics in the years that come.
All of this, of course, against a backdrop of cost unpredictability, which seems to be hitting companies harder than high costs alone.
“Opene,” Abitol said. “The billing is flat, we don’t have any bill shock at the end of the month, and the frontier is the exception.”
