Cloudflare just shared something most companies keep quiet: exactly how their own engineers use AI to write code. The numbers are striking. 93% of their R&D organization, some 3,683 people across 295 teams, actively use AI coding tools built on Cloudflare's own infrastructure. They're pushing 241 billion tokens a month through AI Gateway alone. That's not a pilot program. That's how they work now.
The setup is clever because it's self-reinforcing. Cloudflare built their internal AI stack using the exact same products they sell. AI Gateway handles routing and cost tracking. Workers AI runs inference on open-source models like Kimi K2.5, which they say is 77% cheaper than proprietary alternatives for certain workloads.
A single proxy Worker sits in front of everything. Per-user attribution. Permission enforcement. No client config changes needed. One command authenticates engineers and sets up all their tools. The whole thing runs through Cloudflare Access with Zero Trust authentication baked in.
The impact on output is hard to ignore. Merge requests climbed from roughly 5,600 per week to over 8,700 on a rolling average. The peak week hit 10,952. Nearly double the Q4 baseline. Cloudflare says they've never seen a quarter-to-quarter increase like this.
They also built a Code Mode sandbox using V8 isolates to safely execute agent-generated code. An AI Code Reviewer integrates with CI to keep quality from tanking as volume goes up. MCP (Model Context Protocol) servers connect everything, letting tools like OpenCode and Windsurf tap into internal systems. Cloudflare's new CLI puts the power of these tools directly into developer hands. Same interface. Same permissions. That's the part other companies should copy. Cloudflare's scanner verifies these protocols.