These three things are the same story. First, a sentence: in its annual report filed in February 2026, NVIDIA describes accelerated computing for "computationally intensive workloads such as artificial intelligence, or AI, model training and inference" (NVIDIA Form 10-K, FY2026). Second, a number's home: that language sits over the company's Data Center segment, the line that turned NVIDIA from a graphics company into the default supplier of AI compute. Third, a patent: on June 9, 2026, the company was granted US12651480B2, "Data set generation and augmentation for machine learning models." Connect the dots and you get one business, described from three vantage points.
Follow both the money and the IP. The 10-K language is deliberately broad — "training and inference" — because NVIDIA sells the same accelerators into both halves of the AI workload. That breadth is the bull case management keeps repeating across filing years; the identical "training and inference" framing appears in the FY2025 annual report as well (filed 2025-02-26). The patent estate is where the breadth gets specific: data-set generation and augmentation is the unglamorous plumbing that makes training runs possible, and owning IP there is owning a piece of the workflow that feeds the chips.
The connection isn't decorative. Investors price NVIDIA on Data Center revenue; that revenue depends on customers running ever-larger training and inference jobs; those jobs depend on tooling like the augmentation methods this grant describes. A patent on the tooling and a 10-K on the workload are not two stories — they are the upstream and downstream of the same one.
What the documents don't say is also instructive. The 10-K describes the workload category, not a guarantee of demand; the patent describes a method, not a moat. Neither one tells you whether next year's training budgets hold. But both tell you, concretely, what NVIDIA believes its business is — and they agree.
That agreement is the point of reading a chip company through more than its keynote. The marketing says "AI." The filing says "training and inference." The patent says "data set generation and augmentation." Each is more specific than the last, and each is a primary document you can check.