What Is an NPU? The AI Chip in Every New Laptop and Phone, Explained

What NPU actually stands for, and what it does

NPU stands for Neural Processing Unit. The “neural” part refers to neural networks — the mathematical structure behind most modern AI models. A CPU is built for general-purpose, sequential tasks. A GPU is built for the kind of massively parallel math used in graphics rendering (and, as it turns out, AI training). An NPU is built for one specific job: running an already-trained AI model as fast and efficiently as possible, over and over again, without draining your battery.

That job has a name — inference. Training an AI model (teaching it from scratch) is enormously computation-heavy and almost always happens in a data center on specialized GPU clusters. Inference (using that already-trained model to, say, blur your background on a video call or transcribe your voice) is comparatively light — and it’s the part that NPUs are optimized for.

Why NPUs suddenly appeared in everything

NPUs aren’t new — Apple’s Neural Engine has existed since 2017 in iPhones. What changed is that AI features moved from “nice to have” to central to how operating systems work. Windows’ Copilot+ features, Apple’s on-device Siri intelligence, and Google’s Gemini Nano on Android all depend on running AI models locally rather than in the cloud, for two main reasons:

• Latency: a response generated on-device is instant. A response that has to travel to a server and back adds a noticeable delay.
• Privacy: for AI features that touch sensitive data — your photos, your voice, your screen activity — running the model locally means that data never has to leave your device.

Running these models on a CPU works, but it’s slow and burns battery fast. An NPU can do the same job using a fraction of the power, which is why manufacturers started building them into every mainstream chip rather than treating them as a niche add-on.

How NPU performance is measured — and why the number can be misleading

NPU performance is usually advertised in TOPS (Trillions of Operations Per Second). As of late 2026, the current generation looks roughly like this:

Here’s the catch: TOPS numbers aren’t directly comparable across vendors. Some companies quote the NPU block alone; others quote a combined “platform total” that adds in GPU and CPU AI performance, which produces a much bigger, more marketing-friendly number. The precision used to calculate TOPS also varies (INT8 vs other formats), and none of that accounts for how well-optimized the software layer is. In practice, a chip with a lower TOPS rating but a more mature software stack can outperform a “faster” chip on paper.

The one number worth remembering: Microsoft set 40 TOPS as the minimum NPU performance required for a Windows laptop to qualify as a “Copilot+ PC” — every current flagship chip from Intel, AMD, Qualcomm, and Apple clears that bar comfortably.

What NPUs are actually used for today

• Live captions and real-time translation, processed instantly without an internet connection
• Background blur and framing in video calls, run continuously without taxing the CPU
• On-device image generation and editing (like Photoshop’s Generative Fill running locally, or Windows’ Image Creator)
• Voice assistants that respond without a network round-trip — Siri’s on-device intelligence and similar features on Android and Windows
• Smart search across your files and photos based on content, not just filenames

What NPUs are not good at

NPUs are optimized for one repeated pattern: running a fixed, pre-trained model efficiently. They’re not built for training new AI models — that still requires GPU clusters, usually in a data center. They also don’t replace your GPU for gaming or your CPU for general computing; think of an NPU as a specialist that takes a specific, recurring job off the other two chips’ plates.