The AI PC arrived with a clean pitch: a dedicated neural engine would run AI features locally — privately, instantly, without a subscription — and 40-plus TOPS of NPU performance would be the dividing line between old laptops and new ones. Intel, AMD, Qualcomm and Apple all ship silicon clearing that bar, and PC makers have rebuilt their lineups around it.
Two years in, the hardware has delivered and the defining application has not. That gap is the honest state of the category.
Why it matters
PCs are replaced on multi-year cycles, so a purchase decision made on an AI promise today locks in for half a decade. If on-device AI matures, buyers who skipped the NPU will feel it; if it stalls, the premium paid for TOPS was wasted. The stakes are equally real for the industry, which has bet a replacement cycle on AI features driving upgrades that specs alone no longer do.
How it works
An NPU is a fixed-function accelerator for the matrix math of neural networks, optimized for performance per watt rather than peak speed. Its role in the platform is to run sustained AI workloads — transcription, image processing, small language models — without waking the CPU or GPU, preserving battery life. Windows' Copilot+ features, Apple's on-device intelligence features, and a growing set of third-party apps route work to it through frameworks like ONNX Runtime and Core ML.
The catch is utilization. Most shipping NPU features are conveniences: background blur, live captions, recall-style search, local summarization of small texts. The heavier AI work users actually value — frontier chat models, coding assistants, image generation — still runs in the cloud, because the models that make it good are far too large for a laptop NPU.
Evidence
Market trackers report AI-capable PCs taking the majority share of premium shipments, driven as much by refresh timing as by AI demand. Microsoft's Copilot+ certification set the 40-TOPS floor, and every major silicon vendor now exceeds it. Yet independent reviews consistently find the new machines' most praised qualities are battery life and responsiveness — benefits of the new CPU architectures and process nodes the NPUs happen to ride along with.
Developer adoption is the lagging indicator: the APIs exist on every platform, but the list of third-party applications with meaningful NPU features remains short enough to enumerate.
The competing read
Platform vendors argue this is the normal shape of a platform shift: hardware ships first, developers follow over years, and the killer app is invisible until it isn't. Skeptics counter that the cloud won the last round for structural reasons — model size, update cadence, cross-device sync — and that on-device AI will remain a privacy-and-latency niche rather than the main event. The likeliest outcome is hybrid: small local models handling the constant, personal, latency-sensitive work, with cloud models for everything heavy.
What happens next
Watch local-model quality at the 3-to-13-billion-parameter scale, where an NPU can actually run things well, and watch whether operating systems expose NPU capabilities as stable, documented APIs that developers can build businesses on. The moment a must-have third-party application requires an NPU, the category's promise is kept. Until then, buy the laptop for its battery and its screen.