The Useful Question
English (US)

Do you need an NPU for the AI features you will actually use?

Consider an NPU when a feature you want specifically supports or requires it. Keep a satisfactory PC for cloud AI, investigate compatible hardware for frequent on-device features, and postpone an AI-driven upgrade if the use is still vague. An “AI PC” label or a large TOPS figure cannot tell you which features you will use or how well they will work.

Start with three buying situations

This guide concerns Windows PCs in the US. Before comparing machines, write down one task you want to improve. “AI might be useful someday” leaves far more room for an unnecessary purchase.

Locate the computation

An NPU, or neural processing unit, is specialized hardware for neural-network computation. CPUs handle general-purpose work; GPUs handle many calculations in parallel; NPUs target compatible AI workloads efficiently. These processors can cooperate within a PC. Microsoft's NPU overview explains their roles.

Having an NPU does not automatically put every application on it. Microsoft's developer guidance explicitly says software needs to support the hardware. The application, runtime, model, and drivers determine what can use it.

A request sent to a cloud service is processed on remote computers. Upgrading your local accelerator does not directly accelerate that remote computation. A sluggish browser interface and a slow service response are different problems to investigate.

An application can use compatible CPU, GPU, or NPU processing inside the PC, or send work over the internet to cloud computers. Using the NPU requires software support.
Original editorial illustration
Expand the diagram

Scroll within the diagram horizontally; arrow keys work when focused.

An application can use compatible CPU, GPU, or NPU processing inside the PC, or send work over the internet to cloud computers. Using the NPU requires software support.
Open the diagram at full size

This original schematic separates processing locations. An application can combine them; it is not a measured description of a particular device.

Captioning and translation have different requirements

Windows live captions provide a useful example. Ordinary speech transcription is available with Windows 11 version 22H2 and later. Translation requires a Copilot+ PC running Windows 11 version 24H2 or later. Microsoft's live captions support page distinguishes the two.

For an English-speaking viewer, translating supported Japanese audio into English captions is within the documented language direction. The listed translation outputs are English and Simplified Chinese. That does not establish English-to-Japanese translation support.

If you only want English speech written as English captions, first investigate the existing captioning feature. If you want translation, check the spoken language and desired output separately. The language used by Windows menus does not settle that question.

That distinction can change the purchase entirely. Write the actual need as “Japanese audio to English captions,” rather than simply “AI translation.” It is a small change that makes a requirements page much easier to use.

Camera effects also vary by hardware

Windows Studio Effects includes background effects, framing, and eye-contact adjustments. Microsoft's feature-specific support page lists a subset available on devices with NPUs below 40 TOPS. It separately restricts effects such as Portrait light to Snapdragon Copilot+ PCs.

Check the exact effect, compatible camera, and Windows updates. For a managed work computer, check organizational settings too. If all you need is a blurred background, first see whether your existing meeting application already delivers an acceptable result. Similar-looking features need not use the same implementation or processor.

Copilot+ requirements are a threshold

Treat an “AI PC” sales description as a starting point for reading the configuration. Copilot+ PC is a defined Microsoft category. Beyond the ordinary Windows 11 requirements, the US specifications page lists a compatible processor with a 40+ TOPS NPU, 16 GB of DDR5/LPDDR5 memory, and 256 GB of SSD/UFS storage.

Meeting the NPU threshold alone is insufficient when another requirement is missing. Meeting the complete hardware standard also does not make every feature available in every language.

TOPS expresses trillions of operations per second. It does not count finished pictures or measure answer quality. Actual application behavior also depends on the model, numerical formats, memory, and software support. To compare completion speed, use measurements of the same task with the same application and settings. The NPU figure alone cannot supply that result.

Make a one-application checklist

For each candidate computer, record:

“On-device AI” does not mean every function works offline under every condition. Check initial downloads and sign-in requirements as well as normal operation. If you handle work documents, an NPU label alone also cannot establish where the application sends or retains data.

For a practical comparison, choose one ordinary workload: use the desired camera effect while opening your usual meeting materials, for example. Look at image quality, responsiveness, and battery behavior. Those observations would be more relevant to your purchase than a larger headline number. This article has not performed that test.

If you cannot yet name a worthwhile feature, prioritize the screen, keyboard, weight, battery life, and performance of your normal applications. A supported capability that improves your day is a better buying reason than a vague fear of falling behind.

For the separate question of computation during an AI answer, see what extra AI “thinking” adds.

Sources and limits

Official sources checked October 11, 2026, Japan time. This is not a hands-on performance or battery review. Recheck before purchase or publication and when Windows, supported processors, languages, or feature availability changes.