Laptop and phone makers now advertise NPUs and “AI PCs” as prominently as processors and screens. Behind the marketing is a real shift: more AI features are running directly on devices rather than in the cloud. Here is what an NPU is and whether it matters for you.

What an NPU does

A neural processing unit is a processor designed specifically for the math of neural networks — mostly multiplying and adding large arrays of numbers at low precision. Compared with a CPU or GPU, an NPU performs these operations using far less power. That efficiency is the point: AI features can run continuously without draining the battery or heating the device.

Phones have included NPUs for years, powering camera processing, face unlock, voice recognition and on-device translation. Laptops followed.

AI PCs and “TOPS”

In May 2024, Microsoft introduced Copilot+ PCs, a category of Windows laptops requiring an NPU capable of at least 40 TOPS (trillions of operations per second), along with minimum memory and storage. Chipmakers including Qualcomm, Intel and AMD released processors meeting the requirement. Apple’s Mac and iPhone chips include a “Neural Engine” serving the same purpose and supporting Apple Intelligence features.

TOPS is a rough indicator of peak AI throughput. Comparisons are tricky because vendors may measure at different numerical precisions, and real performance also depends on memory and software support.

What runs on the NPU

Examples of on-device AI features:

  • Video calls: background blur, eye-contact correction, noise suppression.
  • Live captions and translation of audio playing on the device.
  • Image features: upscaling, object removal, generation in some apps.
  • Search and recall of content on the device.
  • Small language models for writing help, summaries and assistants.

Larger AI tasks — the strongest chatbots, long reasoning, video generation — still run in the cloud.

Why on-device AI matters

  • Privacy: data can be processed without leaving the device.
  • Latency: no round trip to a server.
  • Offline use: features work without connectivity.
  • Cost: companies avoid cloud compute for every request.
  • Battery life: efficient NPUs handle background AI tasks with little power.

Do you need an AI PC?

It depends on what you do:

  • If you rely on cloud assistants, any modern computer with a good browser is enough.
  • If you want built-in AI features of the latest operating systems, check their hardware requirements.
  • If you run local language models, memory capacity and GPU power often matter more than the NPU, because many local AI tools primarily use the CPU or GPU.

For most buyers, the NPU is a useful bonus rather than the deciding factor — but it will matter more as software increasingly uses it.

Common misconceptions

  • Buying on TOPS alone. Memory, software support and the features you actually use matter more than peak numbers.
  • Assuming all AI runs locally on an AI PC. Most powerful assistants still run in the cloud; NPUs handle smaller, frequent tasks.
  • Thinking older computers can’t use AI. Cloud AI works on almost any device; some local models run on CPUs and GPUs without an NPU.

Should you buy an AI PC?

If you mostly use cloud AI tools, an NPU adds limited benefit today. It matters more if you want on-device features — live captions, background blur, local transcription or local models — with better battery life. Check which apps actually use the NPU before paying extra.

Related: GPU vs TPU vs NPU.

Key takeaways

  • An NPU is a low-power processor specialized for neural network math.
  • Microsoft’s Copilot+ PCs, introduced in May 2024, require an NPU with at least 40 TOPS.
  • NPUs run features like background blur, live captions and small on-device models.
  • On-device AI improves privacy, latency, offline use and battery life.
  • For running local LLMs, memory and GPU often matter more than the NPU.

Frequently asked questions

What does TOPS stand for?

Trillions of operations per second, a measure of peak AI processing throughput. It is useful for rough comparison but depends on how it is measured.

Do Macs have NPUs?

Apple Silicon chips include a Neural Engine, Apple’s name for its NPU, used by Apple Intelligence and other machine learning features.

Will NPUs replace GPUs for AI?

Not for large workloads. NPUs are optimized for efficiency on smaller models; GPUs remain essential for heavy local AI and for the cloud.