The real scaling of autonomy

The week before I got back to SF, there was a blackout. It shows a weird truth about Level 4 AVs: Yes, there is the long‑tail of perception and planning. Weird edge cases, odd pedestrians, cones in strange places. We all know that story. But the real scaling limit is starting to look a lot more like classic ops math than sci‑fi AI. When something rare or ambiguous happens, cars pause and ask for help. That “help” is a human in a chair, somewhere, juggling multiple vehicles through a remote-confirmation pipeline. In theory, the more data you have, the better the model, the more cars you can deploy. In practice, once the world gets messy, your fleet size starts to become a function of how many teleoperators you can hire, train, and keep awake at 3 am. It is basically an invisible queueing system glued on top of your autonomy stack. And in a citywide failure like the blackout, all those “rare” events suddenly synchronize. So for founders in autonomy and infra: your scaling story is not just about model performance curves. It is also about call center math, queue theory, and human-in-the-loop throughput. The ops slide might be the real safety slide.