Editorial illustration for Revolut Cuts AI Coding Costs by 90% Using Chinese Model After USD 10K Spend
Revolut Cuts AI Coding Costs by 90% Using Chinese Model...
Revolut has slashed its AI coding bill by 90% after swapping Anthropic and OpenAI's commercial models for a self-hosted version of GLM, an open-weight model built by Chinese lab Z.ai. The fintech, valued at $115 billion after a secondary share sale in July, now runs the model on its own NVIDIA GPUs rather than paying per-token rates to US providers. David Abitbol, who heads AI operations at Revolut, laid out the numbers at a technical session during the AI Engineering conference.
The shift caps a yearlong experiment inside the company. Revolut first brought AI tools into its engineering workflow in January 2025, starting with Gemini for internal data access before moving to Anthropic's Claude Code. Costs spiraled fast: some engineers were burning through $1,000 to $10,000 a month with no clear return, prompting Revolut to cap usage.
That left Abitbol's team hunting for a way to give engineers open access to coding tools without the commercial price tag attached. By April 2026, Revolut had started self-hosting open-source models and landed on the GLM family to run its coding workloads, downloading the weights and keeping inference on infrastructure based in the US.
Why this matters
Revolut just gave every CTO a number to point at in budget meetings. A 90% cut in inference spend, plus a 60% drop in coding time across 2,000 engineers, isn't a lab benchmark, it's a production deployment at a 45-million-customer bank. That's the kind of proof point that moves procurement decisions faster than any model leaderboard.
The more interesting signal is what it says about the open-weight ecosystem. Z.ai's GLM running on self-hosted Nvidia GPUs beat the economics of Anthropic and OpenAI badly enough that a regulated fintech bet its engineering workflow on a Chinese lab's weights. For founders watching per-seat AI costs balloon into five figures a month with no clear ROI, as Revolut's own engineers were doing, self-hosting open models stops looking like a hobbyist move and starts looking like a line item worth testing.
What we'd want to see next: independent cost audits, whether GLM's quality holds up outside Revolut's internal benchmarks, and whether other regulated firms follow once compliance teams sign off on Chinese-origin models. That last part isn't settled yet.
Related Reading
Business & Startups Nordic pilot adds Gemini for Education, NotebookLM to boost AI literacy
Business & Startups Kling launches Video O1, all-in-one model with MVL bridge using transformer
Business & Startups DeepSeek Seeks More Capital Weeks After USD 7B Funding Round
From Archives OpenAI raises bug bounty payouts to USD 172,000, matching Google rates
From Archives Coforge BridgeCX Connects Pega, Mastercard, Dynamic Yield for Personalisation
Further Reading
- Revolut says self-hosted GLM models cut its AI coding costs 90% - AI Chat Daily
- Revolut says self-hosted AI cut coding costs 90% - StartupHub.ai
- Revolut Cuts AI Coding Costs by 90% With Self-Hosted GLM Models - Fuller Smith
- Z.ai Review 2026: I Tested It — GLM 5, Features, and Pricing - AI Exposé
- GLM (Z.ai): Best Uses, Limits and Alternatives - Agentarius
