Red Hot Cyber
Cybersecurity, Cybercrime News and Vulnerability Analysis
There comes a moment when an entire industry must stop and ask itself whether the path it is taking is truly the right one. In the world of artificial intelligence, that moment may have arrived.
We at Red Hot Cyber had already said it: the AI race would not remain confined to large American proprietary models, massive cloud infrastructures, and ever‑more expensive subscriptions. The center of gravity is progressively shifting towards China, where many open‑weight models are emerging, usable also in proprietary infrastructures and without necessarily depending on the manufacturers’ cloud services.
The Revolut case is emblematic: one of the most important European fintech players has chosen a Chinese model for its engineers’ daily work, drastically reducing costs. On the other hand, industrial history has taught us something: when an open technology reaches sufficiently competitive performance, it often ends up prevailing over proprietary solutions based on the buy model. It is not only a matter of price, but of control, independence and the freedom to build one’s own technological future.
But back to the headline. Revolut has abandoned Claude Code from Anthropic and Codex from OpenAI as the default programming tools for its roughly 2,000 software engineers.
Now they are using GLM, an open‑weight model from the Chinese artificial‑intelligence lab Z.ai, hosted directly within Revolut’s datacenters. According to Revolut, the spend is now about a tenth of what it previously paid to Anthropic and OpenAI.
This was reported by David Abitbol, head of AI operations and infrastructure at Revolut, when he presented the data in his talk “Ship like a startup, regulated like a bank” at the AI Engineer conference in New York. The details went viral in early November after Trending Topics covered them.
Revolut provides GLM to its engineers using roughly 56 Nvidia B200 GPUs, at an average cost of about $0.30 per million tokens. In a sample of roughly 2,000 engineers, Abitbol reported a 60% reduction in programming time and a 70% reduction in delivery time (from task acceptance to integration) after the switch. These are figures supplied by Revolut itself, and the company’s comparison does not isolate the model change from other factors, such as the removal of token limits.
Revolut initially used Google’s Gemini, then Claude Code with rigid budgets per engineer. Abitbol stated that the initial rollout of Claude Code was costly for a modest gain and that token rationalisation forced engineers to reserve the AI for the most complex programming tasks. With a fixed, self‑hosted invoice, the limits were eliminated. “The bottleneck for engineers is not necessarily programming,” he said, pointing to time lost in email inboxes, Jira and Confluence.
Abitbol acknowledged the “stigma” of using a Chinese open‑source model in a regulated bank. Revolut uses only open‑source weights and runs them on its own infrastructure in the United States, not via Z.ai’s hosted service. Image‑processing tasks are routed to Gemini, while Claude Code and Codex remain available as exceptions for the most complex work. Code agents are run in a test environment with approval rules and human oversight.
The 90% savings exclude cloud, personnel and monitoring costs, and depend on careful optimisation: Abitbol said that the default inference was “terrible” until the team reduced the time needed to generate the first token from about 14 seconds to 1.7 seconds on the same GPUs. Smaller companies, lacking Revolut’s scale and infrastructure team, may not achieve the same result.
One of Europe’s most important tech companies is adopting a Chinese open‑weight model as the standard for the daily programming work of its entire engineering team, no longer just experimentally.
This is a warning for the corporate AI‑lab sector: for routine, high‑volume work, the “good‑enough and cheap” solution is beginning to prevail, while frontier models are reserved for exceptions.
