Autonomy demos show the car. They never show the thermal-limited box in the trunk deciding, sixty times a second, whether it has enough compute left to run every perception model it would like to. That decision is a real constraint, and it is the one this column exists to surface: a self-driving system does not have infinite processing, and running out of it at the wrong moment is a safety problem, not a performance footnote.
The record names it directly. NVIDIA's US12649478B2, “Method to estimate processing rate requirement for safe AV driving to prioritize resource usage” (granted June 9, 2026), is a patent about compute budgeting. Its CPC classes are pure vehicle-control — B60W 60/0015 (autonomous driving control), B60W 50/06 and 50/0205 (improving and diagnosing the control system) — not generic computing. The framing is explicit: estimate how much processing safe driving requires, then prioritize.
The way this actually works: the system continuously estimates the processing rate it needs given the current scene — a quiet highway needs less than a chaotic intersection — and allocates its limited compute accordingly, shedding or deprioritizing lower-value work so the safety-critical tasks always get served in time. It is the same logic an operating system uses to schedule a CPU, applied to the much higher stakes of keeping a two-ton vehicle's perception loop fed.
Why is this the bottleneck nobody prices in? Because compute is finite, power-limited, and generates heat, all inside a car. You cannot simply add another rack. Every additional sensor and every heavier model competes for the same silicon, and the marketing instinct is to keep adding capability. A patent on rate-estimation and prioritization is an admission, in legal language, that the budget is real and has to be actively managed.
There is a quiet honesty in filing on this. A company selling the dream of effortless autonomy has, in its patent estate, a grant whose entire premise is that the onboard computer can be asked for more than it can deliver and must triage. That is the operational reality under the keynote — and it is disclosed, if you read the claims instead of watching the demo.
For anyone evaluating self-driving claims, this is the question to ask after “what can it do?”: “what happens when the compute runs short?” The capability is bounded by the budget, and the budget is bounded by physics. NVIDIA patenting how to manage that budget tells you the constraint is central, not incidental — which is exactly what the demos are built to make you forget.