Show me the line item — and when there isn't one, show me the mechanism. IBM's grant US10671891B2 (“Reducing computational costs of deep reinforcement learning by gated convolutional neural network,” issued 2020-06-02) is a mechanism with no obvious line item, which is precisely why it's worth flagging.
Reinforcement learning — the training paradigm now central to aligning and tuning large models — is notoriously compute-intensive. The patent, assigned to International Business Machines Corporation and classified in CPC G06N 3/04 and 3/08, uses a gated convolutional structure to cut that computational cost. In plain terms, it tries to get the same learning for fewer operations.
The reason a disclosure-first desk cares is that compute cost is the through-line of the entire AI business. Whether a model is trained or served, the bill is operations times price-per-operation. Efficiency IP attacks the first term. It almost never appears as its own revenue or cost line in a 10-K — it disappears into aggregate R&D and infrastructure spend — but it shapes the margins that do appear.
IBM's filings discuss AI as a strategic priority across its software and consulting segments without isolating technique-level economics, which is normal. The patent is the granular record under the aggregate: dated 2020, owned, and aimed at the cost of a training method that has only grown more important since.
The standard caveat applies and is load-bearing: a grant proves invention and ownership, not earnings. We attribute no revenue to it. The reusable point is that AI efficiency IP is real, dated, and held — and that the cost discipline behind model economics was being patented well before the market priced the AI cycle.