What is an AI PC? What the label means, and what it doesn't

“AI PC” is on almost every laptop sticker now, and it means less than it suggests. Here is what the label actually guarantees, and what to look at instead depending on what you want AI to do for you.

The short definition

An AI PC is a computer whose processor includes an NPU (neural processing unit): a small, power-efficient part of the chip designed for the math that AI features use. Intel, AMD, Qualcomm and Apple all build NPUs into their current laptop processors, so most new laptops, and many desktops, qualify.

A Copilot+ PC is Microsoft’s stricter standard. According to Microsoft, it needs:

• an NPU that can perform more than 40 trillion operations per second (40 TOPS),
• at least 16 GB of RAM and 256 GB of storage,
• Windows 11 version 24H2 or newer.

Microsoft lists AMD Ryzen AI 300-series, Intel Core Ultra 200V-series and Qualcomm Snapdragon X chips among those that qualify, and newer generations from all three do too. Many earlier “AI” processors from 2023 and 2024 have NPUs of roughly 10 to 16 TOPS and do not qualify.

What the NPU does for you

The NPU runs AI features that are always on or run often, without draining the battery the way the main processor or graphics chip would. On Copilot+ PCs, Windows uses it for features such as real-time translated captions and on-device image generation, and apps can use it through Windows’ AI programming interfaces.

For most people, that is the whole story of the “AI PC” label: some useful built-in features, and better battery life while using them.

What an AI PC does not tell you

Many people shopping for an “AI PC” actually want to run AI models themselves: chat assistants, coding helpers or document tools that work offline and privately. The label says almost nothing about that, for two reasons:

1. Local AI apps mostly don’t use the NPU. Apps for running language models, such as LM Studio and the llama.cpp engine behind many of them, run models on the graphics processor or the main processor.
2. Memory decides what runs, and how fast. A model has to fit in memory, and it writes about as fast as that memory can be read. In our 21-machine bandwidth study, writing speed tracked memory bandwidth almost perfectly; NPU TOPS don’t enter into it.

So a thin Copilot+ laptop with 16 GB of memory runs small models well, but an older desktop with a gaming graphics card can run the same models several times faster. On our own desktop, an 8 GB RTX 4060 wrote with Qwen3.5 9B at 45.5 tokens per second, against 10.4 on the processor alone.

Which kind of AI computer do you need?

If you want…	Look for
Windows AI features, long battery life	Any Copilot+ PC; choose on screen, keyboard and battery like any laptop
To run small AI models (up to about 9B)	16 GB of memory, or any graphics card with 8 GB
To run mid-size models (20B to 35B)	A 16 to 24 GB graphics card, or 32 GB or more of unified memory
To run the largest open models (70B to 120B)	64 to 128 GB of unified memory: a Mac Studio, a Ryzen AI Max mini PC or a DGX Spark

Our guide to how much memory local AI needs and the Can it run? calculator go deeper, and our laptop, mini PC and desktop guides cover specific machines.