Capex is a promise; the memory wall is part of the fine print. The published application US20220043759A1 (“Techniques for an efficient fabric attached memory,” published 2022-02-10) — which later issued as NVIDIA's grant US11822491B2 (2023-11-21) — is aimed squarely at that wall. The named inventors include Brian Kelleher and Denis Foley, longtime NVIDIA systems architects.
The mechanism is bandwidth, not flops. Modern accelerators can compute faster than memory can feed them, so a growing share of system design is about getting data to the chip across a fabric without stalling it. “Fabric attached memory” is pooled memory accessed over an interconnect; doing it efficiently is what keeps expensive accelerators from sitting idle.
“Fabric Attached Memory (FAM) provides a pool of memory that can be accessed by one or more processors, such as a graphics processing unit(s) (GPU)(s), over a network fabric.”— U.S. Patent Application 2022/0043759 A1 source
For the capex desk this is where the bill quietly grows. The line items hyperscalers disclose are aggregate — “servers, network equipment, and data centers” — and memory and interconnect are buried inside them. But the engineering effort, and the patents behind it, tell you that a meaningful slice of the spend is about feeding chips, not just buying them.
Watching the application mature into an NVIDIA grant is itself the disclosure-relevant event: an idea published in 2021 became an owned, dated grant in 2023. That progression is the kind of thing a patent record shows and a press release doesn't — the bottleneck was identified, claimed, and locked down over time.
The discipline: no filing breaks out “memory fabric” capex, and we attribute no figure to the patent. The defensible claim is that the memory-bandwidth bottleneck is real, expensive, and being engineered — and that the IP behind solving it is dated and now held by NVIDIA.