Capex is a promise; smarter scheduling is one way to extract more from it. The application US20230229916A1 (“Scalable tensor network contraction using reinforcement learning,” published 2023-07-20), with inventors including Brucek Khailany — a name associated with NVIDIA research — uses reinforcement learning to optimize how tensor-network contractions are ordered.
The mechanism is meta: use AI to make AI-adjacent compute cheaper. Tensor-network contraction is an expensive operation whose cost depends heavily on the order operations are performed in; finding good orderings is hard, and learning them with reinforcement learning is a way to cut the compute bill on these workloads automatically.
NVIDIA's Data Center segment is built on making heavy compute efficient, and the company repeatedly patents using machine learning to optimize its own stack. This application is one instance: dated 2023, NVIDIA-linked, aimed at scheduling efficiency that never appears as its own financial line.
Published is not granted, scope is unsettled, and I attach no number — no filing isolates this kind of saving. What it documents is that the “optimize compute with AI” theme, which underpins a lot of NVIDIA's efficiency narrative, was an active 2023 research target with IP behind it.
For the capex desk, the frame is that efficiency compounds across the stack — hardware, kernels, scheduling. Patents that use learning to optimize scheduling are part of why effective capacity exceeds raw spec, and any honest payback analysis has to account for that software layer, not just the chips on the invoice.