Here's what the filing actually says. NVIDIA's FY2026 Form 10-K describes accelerated computing for workloads "such as artificial intelligence, or AI, model training and inference, data analytics, scientific computing, robotics, and 3D graphics." Strip the adjectives and you get the company's own one-sentence theory of its business: it sells general-purpose acceleration, and AI is the largest item on the list.
Why lead a general-reader piece with a workload sentence instead of a revenue figure? Because the sentence is the durable part. Quarterly numbers move; the description of what the chips do is stable across filing years — the FY2025 report (filed 2025-02-26) uses nearly identical "training and inference" language, and the FY2024 report (filed 2024-02-21) does too. When a company repeats the same framing three years running, that's the framing to trust.
The document, not the press release, also clarifies a distinction readers often blur: training versus inference. Training is building the model — a big, intermittent compute job. Inference is running the finished model to answer queries — smaller per request, but continuous and growing with usage. NVIDIA lists both deliberately, because its accelerators sell into each, and the mix between them shapes demand over time.
What the 10-K is careful not to do is promise that demand continues. It describes the workloads its products address; it does not guarantee the buyers keep buying. That separation — between what is disclosed (the product's purpose) and what is inferred (future demand) — is exactly what a careful reader should preserve.
For a non-specialist, the takeaway is simple. If you want to know what NVIDIA is, don't read the ticker. Read the workload sentence in the annual report: training, inference, and a handful of adjacent computing jobs. Everything else in the AI-chip conversation is a gloss on that line.