Revolut GLM Switch Cuts AI Coding Bill 90% for 2,000 Engineers

Revolut has cut its AI coding bill by about 90% after switching roughly 2,000 engineers to Z.ai’s open-weight GLM models running on its own GPUs, according to David Abitbol, who leads AI operations at the London-based fintech. The Revolut GLM setup costs about $0.30 per million tokens, roughly a tenth of what the company had been paying Anthropic and OpenAI.

Abitbol detailed the move in a talk at the AI Engineer conference in New York in October, which The Register reported on November 3. Revolut started self-hosting GLM-5.1 in April and has since moved to GLM-5.3, served from 56 Nvidia B200 GPUs.

Why Revolut Moved to Self-Hosted GLM

Revolut rolled out Gemini to engineers in early 2025 and later tried Claude Code, but kept strict budget limits because some engineers were spending $1,000 to $10,000 a month without clear returns. “We were highly skeptical about the cost of AI,” Abitbol said.

Self-hosting removed those limits, a path that releases like Mistral Large 4 have made more practical for large companies. The bank now pays a flat infrastructure bill instead of metered API fees, and cached tokens are effectively free. “You no longer have to save your tokens for the end of the quarter or the most difficult task,” Abitbol said.

After the GLM rollout, Revolut reports a 60% improvement in coding time and a 70% improvement in lead time, the period from picking up a feature to merging it. Abitbol said the biggest gains came from work outside the editor, such as debugging, email, Jira and Confluence tasks. These are internal figures that do not isolate the model’s effect from other changes.

How the 56 B200 GPUs Are Tuned

Revolut targets less than three seconds to the first token and about 200 tokens per second of output, which Abitbol called enough for engineers who now run several agent tasks in parallel. The talk slides show time to first token falling from 14.26 seconds during the migration to 1.68 seconds on the same fleet, while demand grew 76%.

The gains came from tuning rather than adding hardware: FP4 quantization, separating prefill and decode, and capping context at around 256,000 tokens. “Long context is a scam,” Abitbol said, arguing that a one-million-token window wastes memory for coding work.

Claude and Codex Become the Exception

GLM is text only, so Revolut routes image and UI work to vision models such as Gemini, and engineers can still reach frontier tools like Claude Code and Codex when a task needs them.

The choice of a Chinese model raises security questions, especially after GLM 5.3 built working Firefox exploits in a recent test. Abitbol said the main risk lies in calling APIs hosted in China, not in running open weights on infrastructure Revolut controls, with sandbox rules and logging around every agent.

The case adds pressure on Anthropic, whose revenue hit $17 billion in Q3 largely on coding demand, as open models close the gap at a fraction of the price.