Cloudflare Builds Internal AI Engineering Stack on Its Own Platform, Achieves 93% R&D Adoption
Key Takeaways
- ▸Cloudflare achieved 93% adoption of AI coding tools across its R&D organization, with 3,683 users generating 47.95 million requests monthly
- ▸Developer velocity increased significantly, with weekly merge requests nearly doubling from Q4 baseline to 10,952 in March
- ▸The entire internal AI stack is built on Cloudflare's own shipping products, demonstrating real-world validation of its AI platform capabilities
Summary
Cloudflare has successfully integrated AI into its internal engineering stack by building a comprehensive infrastructure layer that runs entirely on its own platform products. Over the past 11 months, an internal tiger team called iMARS (Internal MCP Agent/Server Rollout Squad) developed the necessary MCP servers, access layers, and AI tooling to enable AI agents to be productive across the organization. The results speak for themselves: 3,683 internal users (60% of the company, 93% of R&D) are actively using AI coding tools, generating 47.95 million AI requests and routing 241.37 billion tokens through AI Gateway in just 30 days.
The internal adoption has had a measurable impact on developer velocity. The 4-week rolling average of merge requests climbed from approximately 5,600 per week to over 8,700, with the week of March 23 hitting 10,952 — nearly double the Q4 baseline. Notably, the entire architecture is built on shipping Cloudflare products including AI Gateway, Workers AI, Durable Objects, Workflows, and the newly announced Sandbox SDK. The company organized its infrastructure into three layers: the platform layer (authentication, routing, inference), the knowledge layer (system understanding), and the enforcement layer (quality control).
- The architecture leverages MCP (Model Context Protocol) servers, AI Gateway for centralized routing and cost control, Workers AI for inference, and newly scaled Workflows for multi-step agent tasks
Editorial Opinion
Cloudflare's approach to building its internal AI engineering stack on its own platform is a smart dogfooding strategy that serves dual purposes: validating product-market fit while demonstrating real-world scalability. The 93% adoption rate and near-doubling of merge request velocity suggest that AI coding assistance, when properly integrated with security and infrastructure controls, can meaningfully improve developer productivity. However, the success may be somewhat self-selecting, given that Cloudflare's engineers have built the tools themselves and likely understand their capabilities and limitations better than external users will.
