AI PC

An AI PC is a notebook or desktop computer with an additional processing unit tailored to AI tasks. This unit is called an NPU, short for “neural processing unit” — a chip that accelerates precisely the kind of computation AI programs need. Regular computers can handle such tasks too, but they consume a lot of power in the process and become slow. The benefit is that programs like automatic subtitles or photo retouching can compute directly on the device. Until now, such requests were mostly sent over the internet to large data centers. The term itself originates from the marketing of chip and PC manufacturers and is not sharply defined.

What manufacturers hope to gain from it

The PC market was boring for years. A computer from 2018 is entirely sufficient for office work and videos, so many people didn’t buy anything new. The AI PC is the industry’s attempt to provide a reason to buy again. Intel, AMD, Qualcomm, Apple, and Microsoft are all pushing this topic simultaneously.

But there are also real advantages for users. When an AI function computes on the device, the data doesn’t leave the computer. Anyone having a meeting transcribed doesn’t send the audio recording to a foreign server. This is a particularly strong argument for companies, law firms, and medical practices.

Then there’s the cost factor. Every request to an AI model in a data center costs the provider money. If the computation runs on the customer’s device, the customer pays for the electricity — which is significantly cheaper for Microsoft or Adobe. That’s why software vendors have a strong interest in AI PCs becoming widespread.

The NPU alongside processor and graphics chip

In an AI PC, three computing units work side by side. The main processor, called the CPU, handles general tasks quickly and flexibly. The graphics chip, the GPU, computes many similar tasks in parallel and is powerful but power-hungry. The NPU is specialized in a single type of computation: a very large number of multiplications and additions with more coarsely stored numbers.

This is exactly the type of computation AI models need. At its core, a neural network consists of millions of such small computational steps. Because the NPU can’t do anything else, it can do this particularly efficiently. For the same task, it often needs only a fraction of the power a GPU would use. For a notebook running on battery, this is the decisive difference.

Performance is measured in TOPS, meaning trillions of operations per second. For its “Copilot+ PC” label, Microsoft requires at least 40 TOPS from the NPU. However, such figures should be read with caution. A high TOPS value says nothing about whether there is even any software that makes use of it.

How to recognize an AI PC in stores

In stores, these devices carry stickers like “Intel Core Ultra,” “AMD Ryzen AI,” “Snapdragon X,” or the Windows label “Copilot+ PC.” Apple doesn’t use the term, but has been building a comparable unit called the “Neural Engine” into its chips since 2017. Modern smartphones have long had such components too, it’s just that nobody notices anymore.

Today, the NPU becomes concretely visible in rather small functions. Video conferences blur the background smoothly without draining the battery. Windows translates spoken language live into subtitles. Image programs remove objects from photos. The most controversial function is called Recall: Windows regularly takes screenshots so that you can search through them later.

In business news, the term mainly appears as a source of hope. Analysts estimate what share of all PCs sold will be AI PCs in the future, and chipmakers' stock prices react accordingly. A common misconception: an AI PC does not run large language models like ChatGPT in their full size. There isn’t enough memory for that. Smaller, stripped-down models run on the device instead.