Strategic Shift in AI Infrastructure
British fintech giant Revolut has reportedly transitioned its artificial intelligence strategy toward self-hosted models, a move aimed at significantly reducing operational expenditures. According to reports circulating on social media platforms such as Reddit, the company has achieved a 10x reduction in AI-related coding costs by moving away from a heavy reliance on third-party providers like Anthropic’s Claude.
The shift follows a period of experimentation where Revolut utilized various frontier models, including Google’s Gemini and Anthropic’s Claude. Industry observers suggest that the company has pivoted toward the GLM model family to handle text-based tasks, while retaining access to Claude specifically for vision-related work.
Cost Optimization and Technical Bottlenecks
The decision to self-host appears to be driven by the high cost of API tokens at scale. While the exact internal financial data has not been independently verified, discussions among developers suggest that for high-volume users, the cost of proprietary APIs can become a significant budgetary burden. By investing in its own infrastructure, Revolut aims to stabilize its AI expenses.
Technical analysis from developer communities indicates that the primary bottleneck for the firm was not the quality of the code generated, but rather the integration of AI into existing workflows, such as debugging and managing documentation within Jira and Confluence.
Some analysts suggest that for specialized tasks like software engineering, the “frontier” capabilities of the most expensive models may not be strictly necessary. Instead, competent developers can often achieve desired results using open-weight models, provided they are paired with the correct “harness” or implementation layer.
The Economic Viability of Self-Hosting
The transition to self-hosting presents a different set of economic considerations. While the reduction in token costs is substantial, the move requires significant upfront investment in hardware and infrastructure.
Critics of the move note that for many companies, maintaining private GPU clusters can become a “money pit” unless the hardware is utilized with high efficiency, potentially running near 24/7. However, proponents argue that coding models do not require the 99.99% uptime demanded by consumer-facing applications, allowing companies to reduce hosting costs by running models only during specific subsets of time.
As open-weight models like the Qwen and Gemma series continue to evolve, the debate over whether to pay for managed API services or invest in proprietary infrastructure is expected to intensify across the fintech sector.