Capex is a promise; utilization is whether you keep it. The application US20230297643A1 (“Non-rectangular matrix computations and data pattern processing using tensor cores,” published 2023-09-21), with inventors including Andrew Kerr and Haicheng Wu — names associated with NVIDIA's compute libraries — targets exactly that. It covers running irregular, non-rectangular matrix math efficiently on tensor cores.
Tensor cores are the dense matrix engines that make AI accelerators fast, but real workloads aren't always neat rectangles. When the math is irregular, naive execution wastes the core's capacity — the chip is busy but not productively. Handling non-rectangular patterns efficiently is about utilization, which is the financial heart of the capex question.
For the capex desk this is the variable behind the headline. Companies disclose how much they spend on AI infrastructure; they don't disclose how efficiently those chips are utilized. A fleet at 60% useful utilization and one at 90% have very different payback profiles on identical spend, and utilization IP like this is one of the levers.
The disclosure caveat is firm: this is a published application, scope unsettled, and no filing attributes a utilization figure or revenue to it. We model nothing off it. The point is that the utilization lever — making expensive silicon do more useful work — was being patented around the NVIDIA stack.
The reusable frame for payback analysis: spend is the numerator everyone watches; utilization is the denominator that decides the return. Patents that improve how accelerators handle real, messy workloads are part of why NVIDIA's effective capacity is more than a spec sheet — and why capex math has to account for software, not just chips.