Here's what actually issued. On February 10, 2026, the patent office granted Microsoft Technology Licensing, LLC US12547872B2, "Machine learning model processing based on perplexity." The named inventors — Bita Darvish Rouhani, Douglas Christopher Burger, and Eric S. Chung — are associated with Microsoft's hardware-aware ML work, which is a useful tell about what the grant is really for. The CPC classifications are G06N 3/042, G06N 3/08, and G06N 3/0495: neural-network methods, core class.
Before scope, the term. Perplexity is a single number describing how surprised a model is by an input — low means predictable, high means the model is struggling. The grant's contribution is to turn that number into a control signal that governs how the input gets processed.
The way to read what this covers is to read it as adaptive compute. The independent claim is directed to processing a machine-learning input in a manner selected on the basis of a perplexity measure — in plain terms, deciding per-input how much computation to spend, keyed to estimated difficulty. Easy inputs take the cheap path; hard inputs get the full treatment. That is the issued idea, and it is narrower and more specific than "adaptive AI," which is exactly why it could issue.
Published is not the standard here — this one is granted, B2, enforceable as of the issue date. But granted is not the same as broad. A claim covers what its language covers. "Processing based on perplexity" is a specific gating signal; it is not a monopoly on the general notion of spending less compute on easy inputs. Read claim by claim, the scope tracks the perplexity mechanism, not the entire genre of efficient inference.
Why would Microsoft prosecute this to grant? Because inference is the cost that scales with usage, and any method that cuts average per-input compute without hurting quality is directly valuable to a company serving models at Microsoft's volume. A granted claim on one such method is a competitive instrument. It does not tell you the technique ships in production — a patent is a method, not a product roadmap — but it tells you where the company chose to plant a flag.
The deflationary note this column always adds: a grant on perplexity-routing is real and issued, and that is genuinely more than most "breakthrough" coverage can say. But it is one method, with claims that narrowed through prosecution to what an examiner would allow. Read US12547872B2 for what it is — an enforceable claim on adaptive compute by input difficulty — and not for what a press release might inflate it into.