Revolut moves internal coding tools to self-hosted AI models to cut expenses

British digital bank Revolut has reduced its AI coding inference expenses by roughly 90 percent over the past year by replacing commercial tools like Anthropic's Claude with self-hosted, open-weight models from Chinese developer Z.ai. David Abitbol, Revolut's infrastructure lead, revealed at recent industry conferences that the firm initially implemented GLM-5.1 before upgrading to GLM-5.3 across a fleet of 56 Nvidia B200 GPUs. Abitbol stated that self-hosting brought costs down to approximately $0.30 per million input and output tokens, about one-tenth the price of commercial alternatives. [1][2]

Before making the transition, Revolut experimented with Google's Gemini in early 2025 and Claude Code in early 2026, but strict budget caps restricted developers because individual usage sometimes climbed between $1,000 and $10,000 each month without clear returns. The GLM-5.1 model features roughly 754 billion parameters, requiring about 380 GB of GPU memory just for model weights and 6 to 26 GB per user for context retention. Alongside the heavy hardware footprint, Revolut limited context windows to around 256,000 tokens rather than one million, allowing it to hit sub-three-second initial response times and 200 tokens per second for its roughly 2,000 engineers. [1][2]

The shift yielded a reported 60 percent rise in coding time productivity and a 70 percent boost in lead time productivity, with engineers using unmetered tokens for debugging and issue tracking. While GLM handles text coding workloads, Revolut still relies on commercial models like Gemini or GPT Luna for visual tasks via a model router. Abitbol noted that self-hosting open-weight models locally eliminates unpredictable end-of-month cloud bills while bypassing the privacy and security issues tied to accessing Chinese API infrastructure directly. [1][2]