Cloudflare Builds Internal AI Engineering Stack on Its Own Platform, Achieving 93% R&D Adoption
Key Takeaways
- ▸Cloudflare achieved 93% adoption of AI coding tools across its R&D organization (3,683 users) using infrastructure built on its own platform products
- ▸The company processed 241.37 billion tokens through AI Gateway and 51.83 billion tokens through Workers AI monthly, routing through multiple LLM providers via a single endpoint
- ▸Developer productivity nearly doubled with merge requests increasing from ~5,600/week to over 8,700/week, demonstrating measurable impact of agentic AI integration
Summary
Cloudflare has successfully integrated AI into its engineering workflow by building an internal AI engineering stack entirely on its own platform products. Over the past 11 months, the company's iMARS (Internal MCP Agent/Server Rollout Squad) team created a comprehensive agentic AI infrastructure that now supports 3,683 internal users across 295 teams, processing 47.95 million AI requests monthly. The infrastructure leverages Cloudflare's existing products including AI Gateway, Workers AI, Zero Trust authentication, and newly announced capabilities like Sandbox and Code Mode.
The impact has been substantial: developer productivity has nearly doubled, with merge requests climbing from approximately 5,600 per week in Q4 to over 8,700 weekly, with a peak of 10,952 in late March. The internal stack architecture comprises three key layers—the platform layer (handling authentication and routing), the knowledge layer (using Backstage and AGENTS.md for system understanding), and the enforcement layer (ensuring code quality through AI Code Reviewer and Engineering Codex). Cloudflare emphasizes that this internal-only infrastructure is built entirely on shipping products, demonstrating the real-world viability of its AI infrastructure offerings.
- The entire internal AI stack runs on shipping Cloudflare products (AI Gateway, Workers AI, Zero Trust, Sandbox, Workflows), validating enterprise readiness of these tools
Editorial Opinion
Cloudflare's approach of dogfooding its own AI infrastructure demonstrates confidence in its platform while providing compelling real-world validation of its product offerings. The 93% adoption rate and near-doubling of developer productivity suggest that well-architected AI engineering stacks can deliver genuine value at scale. This strategy—building internal tooling on shipping products—sets a high bar for transparency and product quality that other infrastructure companies should emulate.
