Revolut cuts AI coding costs 90% with self-hosted GLM

What happened?

Revolut moved its engineers' AI coding onto an open model running on its own GPUs, and its bill dropped to roughly a tenth of what it had paid Anthropic and OpenAI. Revolut engineer David Abitbol presented the details at the AI Engineer conference in New York (talk, transcript). The Register reported on the case on November 3.

The numbers

• About 2,000 engineers use the system for 60% of their coding time, and lead time improved by 70%.
• 56 NVIDIA B200 GPUs run GLM-5.3; the team started with GLM 5.1 in April.
• The blended price is about $0.30 per million tokens. Previously, individual engineers generated $1,000 to $10,000 a month in API costs.
• Time to first token (TTFT) fell from over 10 seconds to 1.6 seconds on the same hardware while demand grew 76%.

How they did it

The key was inference tuning: NVFP4 quantization (FP4 saves about 40% versus FP8), 8-way tensor parallelism, a cache hit rate above 90% and prefill/decode disaggregation. The targets are sub-3-second TTFT and 200 tokens per second, rising to 800 after working hours.

Abitbol called long context "a scam": 256K tokens is plenty for coding. A router sends vision requests to Gemini and OpenAI models, and frontier models are kept as the exception.

In a regulated setting

Revolut is a bank under EU regulation, so its agents run in a human-in-the-loop sandbox with zero data retention and OpenTelemetry-based logging. The GLM weights are hosted outside China, and data stays inside the company's infrastructure.

Why it matters

This is one of the first detailed enterprise cases showing that coding agent costs can drop by an order of magnitude through self-hosting without hurting the developer experience. Smaller teams do not need 56 B200s, but the methods (quantization, caching, routing) work at smaller scale too. Frontier models are best kept for the hard cases.