The quiet trend of 2026 is capable small models running on phones and laptops. For businesses this means AI features without per-request cloud costs, better privacy stories, and hardware refreshes that increasingly assume AI workloads by default.
While the frontier labs traded headlines, small models kept getting shockingly good. Distilled, quantised and tuned, they now handle transcription, summarisation, extraction and classification locally on ordinary hardware.
The business math follows. Cloud model calls are cheap per request and expensive per habit. When routine document handling runs on the laptop itself, the marginal cost drops to zero, and sensitive data never leaves the building, which simplifies privacy conversations considerably.
What to do about it
Notice it at refresh time: the AI-capable hardware premium keeps shrinking, and mainstream business machines increasingly carry NPUs as standard. Buying for the next four years means buying for local AI whether you plan to or not.
Architecturally, prefer tools that can degrade gracefully to local processing for routine work and reach for frontier cloud models only where the task truly needs them. Cost and resilience both improve.