Revolut's Shift to Self-Hosted AI Signals Maturing Platform Strategy for Cost and Control

Revolut, the global financial technology company, has transitioned approximately 2,000 of its engineers from commercial AI coding assistants like Anthropic's Claude Code and OpenAI's Codex to self-hosted GLM models. This strategic move, detailed by Revolut's Ops and Infrastructure leader David Abitbol, has reportedly slashed AI coding costs by around 90%. The shift was driven by economic shakiness with early commercial AI rollouts, where engineers were consuming large token budgets without clear ROI, leading to usage limits that inadvertently stifled broader AI adoption. This development is significant for platform engineering practitioners because it demonstrates a tangible commitment to building internal AI capabilities as a core platform offering. The primary motivation extends beyond mere cost reduction; it's about empowering developers by removing the cognitive load and friction associated with external AI services, such as unpredictable billing and token scarcity. By self-hosting, Revolut aims to provide its engineers with unconstrained access to AI, fostering wider integration into daily workflows beyond just coding. This allows engineers to delegate more tasks to AI, addressing bottlenecks not just in code generation but also in surrounding activities like managing inboxes and Jira tickets. This trend aligns with the broader evolution of platform engineering, where internal developer platforms (IDPs) are becoming central to delivering not just infrastructure but also advanced capabilities like AI as a self-service product. As organizations mature in their AI adoption, the need for governance, cost control, and seamless integration into existing toolchains becomes paramount. Platform teams are increasingly responsible for providing 'AI-ready' infrastructure, encompassing GPU orchestration, model versioning, inference optimization, and robust MLOps pipelines. The move to self-hosting reflects a desire to internalize the operational complexities of AI, much like how platform teams have internalized cloud infrastructure management to provide golden paths for application development. In practice, this means platform engineers should be evaluating the trade-offs between consuming managed AI services and building in-house solutions. Factors to consider include the scale of AI usage, cost predictability, data privacy requirements, and the level of customization needed. Revolut's experience highlights the importance of tuning serving systems before scaling GPU investments; they achieved significant performance improvements (76% higher requests per second, 88% cut in time to first token) by optimizing their existing setup rather than immediately adding more hardware. Practitioners should focus on building robust internal platforms that can efficiently serve AI models, manage resource allocation, and provide clear observability into AI consumption and performance. This also implies a growing need for platform teams to develop expertise in AI infrastructure, including model deployment, inference optimization, and GPU management, to effectively support their organization's AI initiatives.