For decades, security companies protected infrastructure and applications from the Internet by stacking boxes, appliances, gateways, and inspection tools around them. The assumption was simple: build the infrastructure first, layer security on top of it after. Traffic has to reach the edge of that infrastructure before anything even looks at it — which means the attacker is already close by the time anything gets stopped.

Cloudflare asked a different question: what if the Internet itself could be that security layer?

That question goes back to 2004, and a much smaller one: where does email spam actually come from? To answer it, the founders built Project Honey Pot — a distributed system that let any website owner plant tracking traps for spammers and malicious bots, mapping their behavior across the internet in real time. Over five years, thousands of websites in 185 countries joined. The dataset grew rapidly, and users kept pushing for more: don’t just track the bad guys, stop them.

That question is the origin story. And here’s what it actually took to answer it.

Lee Holloway didn’t build another standard web proxy. He built something the internet had never seen: a globally distributed reverse proxy layer, running the exact same software stack on every machine, everywhere, simultaneously. The physical infrastructure was unremarkable — commodity x86 servers, sitting in colocation facilities around the world, nothing exotic. The radical part was what the software was designed to do.

Instead of sending traffic to one system for caching, another for security, and another for routing, Lee built a single unified pipeline — a request arrives, gets parsed, hits security logic, gets routed, and gets served, all inside the same system, in the same pass. Combine that with anycast routing, where every Cloudflare location shares the same IP address and the internet automatically routes each user to the nearest one, and something remarkable falls out of it: any Cloudflare server, anywhere on Earth, can handle any request, for any customer.

Why didn’t everyone build it this way? Because it is brutally hard. It required writing high-performance networking code so security wouldn’t slow anything down. It required solving distributed systems problems most companies avoid entirely — pushing policy changes globally in seconds, keeping every location consistent, failing over gracefully when parts of the network go down. And it required walking away from the business model most networking companies were built on: selling high-margin hardware appliances. Cloudflare’s entire bet was that if you own the network, you don’t need to sell boxes. They didn’t build services on top of a proxy — they built a network.

And the numbers now prove the bet paid off. Roughly a quarter of the world’s Internet traffic runs through this network today. It’s built to sit within 50 milliseconds of nearly every connected person on Earth — more than 337 cities, over 125 countries — with AI inference running from over 210 of those locations, offering access to more than 350 different models, and 80% of the top 50 generative AI companies running on Cloudflare.

But the numbers were never the real story. The architecture is. Because that same architecture — one pipeline, already in the path, before the request reaches anything important — is exactly why Zero Trust makes sense, why API security makes sense, why AI security makes sense. Not tacked on later. Already there.

Which brings us to why this matters right now more than ever. AI agents are not like traditional software. Traditional software runs in predictable locations, on predictable schedules, talking to known endpoints. AI agents are autonomous — they make decisions, call APIs, spin up processes, talk to other agents, constantly, globally, simultaneously, at a scale that was unthinkable five years ago. That kind of traffic needs infrastructure that is globally distributed, low-latency, secure by default, and instantly available — no servers to provision, no regions to choose.

Cloudflare has been building exactly that infrastructure for fifteen years, without knowing AI agents would ever need it. That foundation was not built for AI. But building for the hardest problems on the internet — global scale, millisecond latency, consistent security everywhere, no boxes — turns out to be exactly what AI needs.

Cloudflare didn’t predict AI. They just built the right network for it. And AI arrived.