R&D capitalized is R&D hidden, and NVIDIA's autonomy R&D was hiding in plain sight by 2021. NVIDIA Corporation's August 3, 2021 grant US11080590B2 claims stereo depth estimation using deep neural networks — perception software, not a chip.
Read the segment, not the GPU spec. The CPC tags are pure machine vision: G06N 3/0454 and G06N 3/0481 network architectures, G06N 3/063 and G06N 3/084 training, G06T 7/593 stereo depth, G01S 17/86 ranging. NVIDIA was patenting the model that runs on its hardware, capturing margin at a second layer.
The decoder problem is that this attach was invisible in the financials for years. NVIDIA's automotive line was small, and the perception-software value embedded in its Drive platform did not surface as a separate revenue stream. The patents were ahead of the income statement.
For a fundamentals reader, that lag is the opportunity and the trap. The IP told you in 2021 where NVIDIA intended to capture autonomy margin; the segment disclosure would not confirm it for years. Read the patents to see the strategy before the numbers ratify it.
The honest limit: a depth-estimation patent is a method, not a market share. It does not tell you how many autonomy programs license NVIDIA's perception stack or at what price. It tells you NVIDIA was deliberately building above the chip.
The takeaway for the money desk: NVIDIA's autonomy thesis is a software-attach thesis, and the patents lead the revenue. When the dominant compute supplier also owns the perception model, every customer's stack is partly NVIDIA's.