35 Billion Dollars and Still Late

The FAA is investing $875 million in an AI system to manage air traffic congestion, starting with Washington's airspace. This is not a new problem being solved — it is a known problem being solved the same way it has already failed.

In 1995, the FAA committed to NextGen, a satellite-based air traffic management system that would revolutionize how planes navigate American airspace. The promise was specific: implementation by 2015, with meaningful congestion reduction across major hubs. The result: $35 billion spent, every deadline missed by five to ten years. Congestion at Atlanta, Dallas, and Chicago continuing to worsen throughout the rollout.

The failure was not technical incompetence. NextGen worked — planes can be routed more efficiently when pilots have better positioning data. The failure was architectural. Congestion at major hubs is not a routing problem; it is a slot problem. Atlanta's Hartsfield-Jackson has five runways. During peak hours, every one receives planes separated by 60-90 seconds. No algorithm makes those five runways accept six planes per minute. Better routing merely concentrates demand at the same bottleneck faster.

The FAA learned nothing. The DC pilot program treats the new AI system as if optimization could solve what only infrastructure or flight reduction can solve. It will likely work exactly as NextGen worked: the software improves, congestion metrics momentarily flatten. Two years in someone discovers that planes still queue on the tarmac while the system hums at maximum efficiency. Success metrics quietly shift. The investment becomes a sunk cost that justified its own continuation.

This pattern. Mistaking a constraint problem for a coordination problem. Shapes how we solve almost everything. We optimize the wrong variable, achieve the optimization, and call it progress while the actual bottleneck remains untouched. The reader building a creative practice often does this: perfecting the studio while the real problem is permission to begin. The manager does it redesigning workflows while the constraint is that there are too many concurrent priorities. Naming what cannot be algorithmed away is rarer than building something that runs smoothly against it.