Most autonomy patents claim a capability; this one claims a way to know whether the car can afford to run it. NVIDIA's grant US12649478B2, "Method to estimate processing rate requirement for safe AV driving to prioritize resource usage" (issued June 9, 2026), is classified under B60W 60/0015 (autonomous operation), B60W 50/06 (monitoring or diagnosing the control system), and B60W 50/035 (responding to abnormal conditions) — the functional-safety codes, not the perception ones.
The problem is real and underdiscussed. An autonomous driving stack runs dozens of compute-hungry tasks at once — perception, prediction, planning, mapping — on finite on-board silicon. At highway speed in dense traffic, the safe-processing requirement spikes; in an empty parking lot it falls. If the compute budget is exceeded, something has to give, and on a moving vehicle "something gives" cannot mean a dropped frame at the wrong moment. The patent claims estimating how much processing is required to drive safely under the current conditions, and then prioritizing resources to guarantee it.
This is an enabling claim of a particular, unglamorous kind. It does not make the car perceive better or plan smoother; it makes the system honest about its own compute headroom and forces graceful prioritization when that headroom shrinks. The inventors are associated with NVIDIA's safety and reliability research, which fits the classifications exactly — this is about the dependability of the compute platform, the thing NVIDIA actually sells into cars.
What is the defensible scope? The novelty is the method of estimating the safe processing-rate requirement and using it to prioritize resources — not the idea of resource scheduling in general, which is ancient in computing. The independent claim ties the estimation to AV safety specifically, which is what distinguishes it from generic real-time scheduling prior art. Reading it as "NVIDIA patented allocating compute" would miss the safety-coupling that is the inventive core.
Why this is a strategic filing for a chip company: it makes the case that safe autonomy is not just about raw TOPS but about provably allocating them. That argument sells more capable platforms — and a patent on the allocation method is a piece of IP that travels with the silicon. For NVIDIA, owning the safe-compute-budgeting method strengthens the platform story it tells automakers.
For the B60W 50 functional-safety subclass, grants like this are the quiet frontier. The headline autonomy patents are about seeing and planning; the durable ones are increasingly about whether the system can certify that it had enough resources to do either safely. That is where this grant sits, and it is precisely scoped to that question.