A neural processing unit is a small block of silicon built to do one job well: the matrix arithmetic that machine-learning models spend most of their time on. It is not a faster computer. It is a cheaper way to run one expensive kind of work, which is why every laptop chip now advertises one and almost nobody can say what theirs actually does.

What an NPU actually does

A model running on your machine is mostly a long sequence of multiplications between grids of numbers. A general-purpose processor can do that arithmetic, but it spends much of its energy moving values between caches rather than calculating with them. An NPU arranges the arithmetic so the numbers stay close together, in a fixed pattern that suits the shape of the problem.

That is the whole trick. The trade is flexibility. A graphics processor can be reprogrammed for a new kind of model. An NPU is closer to a dedicated appliance: efficient at what it was designed for, useless at everything else.

Why 40 TOPS became the number

In May 2024 Microsoft introduced Copilot+ PCs and defined them by a single figure: a neural processing unit capable of more than 40 trillion operations per second. The company's developer documentation still says many Windows AI features require an NPU that can run at 40+ TOPS.

That requirement did two things at once. It gave laptop makers a target, and it turned a technical specification into a marketing line. TOPS measures how many operations a chip can perform in a second, which is a real number and a poor proxy for whether a feature will feel good. It says nothing about memory bandwidth, model support or the quality of the software on top.

What changed in 2026

Two years on, the hardware finally caught up. At CES 2026, Intel's Core Ultra Series 3 chips carried a 50 TOPS NPU, AMD's Ryzen AI 400 series reached 60 TOPS, and Qualcomm's Snapdragon X2 Elite and X2 Plus were quoted at 80 TOPS. Copilot+ capable machines stopped being unusual.

The software strategy moved in the other direction. Microsoft's Windows AI Foundry, which replaces the Copilot Runtime that required an NPU, lets developers run inference on a GPU, a CPU or an NPU. Microsoft also said that every Windows 11 PC would be an AI PC, not only the Copilot+ ones. As PCWorld put it in January, laptop makers spent two years chasing a requirement that Microsoft then stopped insisting on.

The NPU did not disappear. It stopped being the entry ticket.

What it means when you buy

For most people the benefit is power, not speed. Background blur on a video call, live transcription, photo processing and on-device summarising can all run on an NPU for a fraction of the energy a GPU would use, which shows up as battery life rather than a higher benchmark score.

The features that exist today are modest. Windows Studio Effects, image generation in the Photos app, Recall for searching your own activity, and Click to Do for acting on text on screen. None of them justifies a purchase on its own, and several work better with a decent GPU.

A sensible reading of a spec sheet: an NPU is a signal that the chip is current, not a promise that the laptop will feel quicker. Memory, storage, screen and keyboard still decide whether you enjoy using it.

The limits

Three limits are worth knowing. An NPU only accelerates models its vendor has prepared it for, so a new architecture can arrive without support. It does not train models, which remains a GPU job. And a high TOPS figure can sit next to a small memory pool, in which case the chip spends its time waiting rather than calculating.

There is also a structural point. Microsoft required NPUs, then built a toolchain that does not depend on them. Hardware requirements set by one vendor's product strategy are not a law of nature, and buyers should treat them as marketing until the software proves otherwise.

FAQ

Do I need an NPU to run AI features? No. A GPU can do the same arithmetic and often does it faster. The NPU exists because it uses less power for the same result.

Is an NPU the same as a graphics card? No. A graphics card is general purpose and programmable. An NPU is a smaller, fixed-function block with a narrow set of operations.

Will an NPU make my laptop faster? Only for tasks that use it. Web browsing, spreadsheets and video editing will not notice it exists.

Do higher TOPS mean better AI? Not reliably. TOPS describes throughput on a specific kind of operation. Feature quality depends on the model, the memory and the software around it.

The short version

An NPU is specialised hardware for the maths behind machine learning, designed to save power rather than win benchmarks. It is useful, it is no longer rare, and it is not the reason to choose one laptop over another.