On the infrastructure that actually matters for agentic AI, and why nobody wants to hear about it.
On June 30, 1956, two airliners collided over the Grand Canyon and killed everyone aboard. Both aircraft were in uncontrolled airspace. Controllers watched both trajectories converging on radar. They could see. They could advise. They could not direct. The system was structurally capable of observing the problem and structurally incapable of resolving it.
The airframes were fine. The engines were fine. The navigation worked. The pilots were competent. Two years later it happened again — this time inside controlled airspace, on a published federal airway — because civil routes and military training corridors overlapped without a binding protocol to deconflict them. Two capable actors, each following their respective procedures, composed into an outcome the system had no graceful way to absorb.
The reform was not a better airplane. It was the Federal Aviation Act of 1958: a single civil authority with rulemaking and operational control over the entire national airspace. Radar continued to improve. Navigation advanced. But what changed the system was a control plane — a layer that could enforce separation rather than suggest it; that could treat the sky as a single programmable surface rather than a patchwork of overlapping jurisdictions where “control” meant awareness without authority.
The planes got smarter. More importantly, the sky did.
This is where agents are now.
There is an infrastructure version of the Prometheus myth that plays out every decade. The primitive captures imagination. The coordination layer captures the return. The audience applauds the fire. The money is in the fire code.
VMware built the hypervisor — the ability to run multiple operating systems on shared hardware, to treat compute as fungible. Within a decade, the hypervisor was commoditized into every OS kernel. What wasn’t commoditized: vMotion, vCenter, the ability to treat a datacenter as a single programmable surface. The primitive virtualized the machine. The control plane virtualized the fleet. By the time Hyper-V and KVM had eroded the hypervisor margin, VMware had already ascended the stack.
Docker gave us the container. Docker raised $275 million on the premise that owning the container meant owning the infrastructure layer. Docker crossed $100 million ARR in 2023. A good business. Not the business they raised for. Docker raised as a platform. It became a tool. By 2020, Docker Hub had accumulated over 130 billion image pulls. The ecosystem exploded. And Docker sold its enterprise division to Mirantis. Volume was enormous. Value capture was elusive.
Because volume without orchestration is noise.
Kubernetes became the platform. Google open-sourced it in 2014, donated it to the CNCF in 2015. The specification is free. The reference implementation is free. Anyone can run it. Anyone can fork it. The intellectual property is, in the most literal sense, given away. And yet Kubernetes created hundreds of billions of dollars in value — accruing to the managed services (GKE, EKS, AKS), the observability vendors, the security vendors, the consulting firms. The container market hit $5.85 billion in 2024, trajectory toward $31.50 billion by 2030. Nearly all of it flows to compute providers, not to Docker.
Each primitive generates a coordination problem that the next layer solves. Virtualization created VM sprawl; the fleet manager collects the rent. Containers created deployment chaos; the orchestrator collects the rent. The primitive is essential. Essential does not mean defensible.