Revolut, Europe's biggest tech startup, cut its AI coding inference costs by 90%. The fintech giant achieved this by ditching Anthropic and OpenAI for a self-hosted model from Chinese AI lab Z.ai. The move also boosted engineering lead times by 70% and coding time by 60% for its 2,000 engineers.
How We Got Here
Revolut began experimenting with AI coding tools in January 2025, but faced monthly engineer spending of $1,000-$10,000 on commercial models like Claude Code. Concerned about spiraling costs, the company imposed strict usage budgets before moving to self-host open-source models by April 2026.
The Numbers
- Blended cost of tokens dropped to 30 cents per million, a one-tenth reduction from previous Anthropic and OpenAI invoices.
- The self-hosted solution runs the GLM-5.3 model on 56 NVIDIA B200 GPUs, serving approximately 2,000 engineers.
- Revolut found removing token constraints let engineers spend less time searching documentation or coordinating tasks.
- Inference performance increased by 88% between April and August, with time to first token falling from over 10 seconds to 1.6 seconds.
- Revolut uses US-based infrastructure for inference, downloading model weights instead of sending requests to Chinese-hosted endpoints.
What Happens Next
🇮🇳 Why This Matters for India
For Bangalore and Hyderabad's dev-tools founders building AI-powered assistants, Revolut's success proves open-source models can deliver enterprise-grade performance and critical cost savings.
The Take
Revolut's decisive shift to self-hosting GLM suggests cost was a symptom, not the core problem; control over data, latency, and customisation was the deeper driver. Expect more heavily regulated fintechs to follow this path, prioritizing sovereignty over convenience within 18 months.
Source:
Analytics India Mag ↗