The fallback claim reveals the anticipated failure. NVIDIA's grant US12649478B2, "Method to estimate processing rate requirement for safe AV driving to prioritize resource usage" (issued June 9, 2026; four inventors), claims estimating how much processing throughput safe driving requires in a given situation, then prioritizing resource usage to meet it. Read past the efficiency framing and the real subject is failure: what the vehicle does when it cannot compute fast enough to drive safely.

The CPC stack confirms it. This grant is unusually thick with safety and diagnostics classes — B60W 50/06 (fail-safe operation), B60W 50/0205 and B60W 50/035 (fault detection and response), and B60W 60/0015 (autonomous operation accounting for safety), plus B60W 2050/143. That cluster is the signature of a claim built around degraded operation. The inventors weren't optimizing for speed in the happy path; they were fencing the policy for when resources fall short of the safe-driving requirement.

“An estimation model utilizes simulations of an autonomous vehicle and objects detected near the automated vehicle to develop estimates of tolerable frame processing latency to develop real world frame processing latency estimates for similar driving conditions.”— U.S. Patent No. 12,649,478 source

Why this is the interesting kind of claim: the honesty of an autonomy stack lives in its fallback logic. Any system performs when compute is abundant. The differentiator — and the litigation-relevant scope — is the minimal-risk behavior when it isn't. By claiming a method that estimates the required processing rate and prioritizes to hit it, NVIDIA is fencing the resource-triage decision that sits directly upstream of a minimal-risk maneuver. That decision is where a safety case is actually made or lost.

For claim-scope analysis, the prioritization element is where the metes and bounds live. The claim isn't a monopoly on "safe driving needs compute" — that's unpatentable common sense. It's a method for estimating a specific throughput requirement and allocating against it. A competitor who guarantees safety by over-provisioning compute, rather than by dynamically estimating and prioritizing, may well sit outside the fence. The novelty is the estimate-and-triage loop, not the goal of safety.

Caveats, in the desk's voice. Safety-method claims invite close prior-art scrutiny because the field is crowded; the examined B2 grant means it cleared that bar, but the prosecution history is where you'd see how far it was narrowed. A method claim describes a procedure, not a certification — nothing here proves a vehicle is safe, only that NVIDIA owns a way to manage compute toward safety. And CPC density signals intent, not scope; claim 1 is the authority.

For the autonomy-software beat, treat this as a fallback-logic filing wearing a resource-management title. The cluster of fault and fail-safe CPC codes tells you what NVIDIA most fears — running out of compute mid-decision — and what it moved to protect. Verify the prioritization limitation in the patent record, and file it where it belongs: with the minimal-risk-condition IP that defines real autonomy.