Spend versus return, again — and Intel's grant US10963783B2 (“Technologies for optimized machine learning training,” issued 2021-03-30) is on the return side of that equation. Assigned to Intel Corporation and classified CPC G06N 3/08, it covers optimizing the training process itself so the same hardware does more learning.
Throughput is the quiet variable in every capex argument. The headline number is how much silicon a company bought; the number that determines payback is how much model that silicon actually trains per dollar and per watt. Training-optimization IP attacks that ratio directly, and it does so in a way that never surfaces as its own financial line.
Intel's position in the AI buildout is as a supplier and as an operator of its own compute, and its disclosures discuss AI across data-center and product lines without isolating training-efficiency economics. The patent is the technique-level record under the aggregate — dated 2021, owned, aimed at the throughput question.
For the infrastructure desk, the reusable frame is denominators over numerators. Two companies can buy identical fleets and get very different effective capacity depending on how efficiently they train. Patents like this are where that difference is engineered, which is why they belong in any honest read of AI payback.
The discipline holds: a grant is invention and ownership, not a disclosed return. We model nothing off it. The claim is that the throughput lever — the thing that turns capex into capability — was being patented at Intel in 2021, on the record and dated.