Revolut has reportedly put Claude Code aside in favor of a self-hosted Chinese model, with a stated cost reduction of 90%. The news, which emerged from a developer discussion, is more than a tool swap: it shifts the economic calculation from per-seat or per-token fees to fixed infrastructure and operations costs.

For a financial company, proprietary assistant API pricing can quickly become a significant line item. Moving to a self-hosted model does not eliminate costs; it transforms them. You need GPUs, VRAM, the skills to serve the model for inference, and processes for updates and evaluation. The 90% saving should therefore be read as a reduction in the variable component, not necessarily an equivalent TCO cut. If infrastructure already exists or is shared with other workloads, the net saving can remain high; if it requires new dedicated nodes, the comparison changes.

The structural signal is clear: Chinese open or available-weight models are becoming competitive not only in generic benchmarks but also in workflows that companies once considered the preserve of Western vendors. That shifts incentives. For proprietary assistant vendors, the competitor is no longer just another SaaS provider; it is the internal team that can download a model, adapt it to its own pipelines, and treat code as an asset under its control. For hardware makers, demand grows for machines suited to local hosting. For regulators, data residency and access control become easier to demonstrate when the model never leaves the corporate perimeter.

Of course, self-hosting a Chinese model introduces new complexities: licenses, weight provenance, security audits, and dependence on a tooling ecosystem that must be maintained. It is not a shortcut for everyone. But if a company like Revolut chooses this path and reports it as a 90% saving, other organizations with sufficient resources are likely to run the same calculation. At that point, the game will no longer be played only on completion quality, but on the ability to integrate self-hosted models into regulated environments without turning them into hidden operational costs.

For those evaluating on-premise deployment, AI-RADAR offers analytical frameworks at /llm-onpremise to weigh these trade-offs.