Sticker Shock: The AI Badge on Your Laptop Says Little About What Its Chip Does

A neural processing unit, or NPU, is a chip built for the arithmetic AI models run on, and makers now put one in laptops and robots alike.

Intel said on August 25, 2026, that its new Wildcat Lake Core Series 3 processors include an NPU rated at up to 17 trillion operations per second, sitting beside CPU cores and integrated graphics ^[3]. The chip maker NXP describes a robot platform that pairs a main processor with a separate NPU, called the Ara240, for vision and generative-AI work done on the machine itself ^[4].

The pattern is clear enough. The AI chip is no longer only a room full of servers. It is a small block of silicon inside the thing you carry or the thing that rolls across a warehouse floor.

What are the three chips in the box?

Start with the two you already know. The CPU, or central processing unit, is the generalist. It handles whatever a computer asks of it, from opening a file to running a spreadsheet ^[5].

The GPU, or graphics processing unit, is a team. It was built for graphics, which means doing a huge number of similar calculations at the same moment, and that talent turns out to suit many other jobs ^[5].

The NPU is the specialist. It is designed for the calculations that neural networks use, the same sums repeated again and again ^[5].

Think of a kitchen. The CPU is the cook who can make anything. The GPU is a row of cooks all chopping onions at once. The NPU is the one person who does nothing but make one dish, all day, and is very good at it.

The specialist does not replace the others. Microsoft's developer guidance says an NPU is not a substitute for the CPU and GPU, and that a system can give each kind of work to the processor best suited to it ^[1].

What is an AI model actually doing?

A model is a recipe that was learned from examples. Making the recipe is called training. It adjusts the model using large amounts of data, and it is the heavy, expensive part ^[5].

Using the recipe on something new is called inference. You show the model a photo, or type a question, and it produces an answer ^[5]. NPUs are aimed mostly at this second job, done efficiently on the device, and not at training giant models ^[5].

Inside, a neural network is layers of calculations that turn an input, such as an image, into an output, such as a label for what is in it. Many of those calculations are matrix operations, which are grids of numbers multiplied together over and over ^[5]. Repetitive, predictable sums are exactly what a specialist can be built for.

Why add a third chip at all?

Because of the battery. A laptop or a robot has a fixed amount of energy and nowhere to plug in mid-task.

Microsoft gives a concrete case. Its Phi Silica is a small language model, meaning a text-writing AI compact enough to run on the device itself. It can run on Copilot+ PCs using the NPU, and on some other Windows PCs using the GPU. Microsoft says the quality of the output is the same either way, but speed, power use and responsiveness can differ ^[2].

The company says the NPU route is tuned for lower delay and lower power. Its comparison says running the model on a GPU can use more power, compete with other graphics work and put pressure on graphics memory ^[2]. So an NPU does not make a model smarter. It lets the model run without hogging the chip you also need for your video call, and without draining the cell as fast.

There is a second benefit. When the work happens on the device, the input does not have to be sent to a cloud service, and the feature can work without a connection ^[2][4][5]. That matters for a laptop on a plane. It matters more for a robot.

NXP says its demonstration robot does autonomous navigation and detects hazards such as fire and oil spills, using the local NPU for vision and generative-AI workloads ^[4]. A machine that has to spot a fire does not want its eyes to depend on a server somewhere else answering in time. That is NXP's own account of its demo, not independent testing.

So what does TOPS mean?

TOPS stands for trillions of operations per second. It is the number printed most proudly on NPU spec sheets.

Microsoft says Copilot+ PC features require an NPU capable of at least 40 TOPS ^[1]. That is 40 trillion simple operations in a second. To feel the size, picture 40 million people each doing a million sums in that one second.

Two cautions come with the number. First, the 40 is a bar for one product category, not a universal measure of whether every AI task will run well ^[1]. Second, TOPS is a peak figure. It says what the chip could do at its best, not how fast your app will feel or how long your battery will last.

Here is a pointed comparison. Intel's new NPU is rated at up to 17 TOPS, which is below Microsoft's 40 TOPS line for Copilot+ ^{[3] [1]}. That does not make Intel's chip useless. It shows that one number can mean a gate for one program and nothing at all for another.

There is a further wrinkle. Microsoft says models may have to be converted to lower-precision number formats, a process called quantization, to run on some NPUs ^[1].

Think of rounding 3.14159 to 3.14. You save effort and lose a little detail.

How a chip maker counts its operations, at what precision and under what test conditions, can change the headline number, and how comparable TOPS are across companies is still an open question.

Does having an NPU mean the feature runs on it?

No, and this is the point where labels and reality part.

Microsoft says NPU performance depends on software and models being set up to use the chip ^[1]. Its Windows software can pick an available accelerator, meaning a chip such as the NPU or GPU that can take on the work, and it can fall back to the GPU or the CPU if the preferred chip is not available or fails ^[1]. Microsoft also notes that some AI tasks can run on a GPU instead ^[1].

So a laptop can carry an NPU, show an AI badge on the lid, and still run a given feature on the CPU. The chip is the engine. Whether anyone has connected it to the wheels is a software question.

How do you read a spec sheet?

Four questions will do most of the work.

First, what does the TOPS figure measure? Ask whether it is a peak, at what number precision, and whether it covers the NPU alone or the whole chip. Chip makers do not always say, and the figures are not guaranteed to be comparable.

Second, is it on the right list? If a maker says a feature needs a 40 TOPS NPU, a 17 TOPS chip is outside that program, whatever else it can do ^{[1] [3]}.

Third, does the app you want support that NPU? The software has to send the work to the chip, and the model may need converting before it can run there ^[1]. "AI-ready" on a box says nothing about your apps.

Fourth, has anyone measured the feature running on the NPU? A claim of lower power is a claim until a test shows the feature using the NPU and shows the battery numbers.

The claims above come from the makers themselves. Microsoft, Intel and NXP describe their own products, and Rapidus, which wrote the plain explanation of the three chip types, is a semiconductor company too ^[1][3][4][5]. None of them is a test lab.

What is not known yet?

Independent measurements are the gap. Real-world battery, speed and power comparisons for laptop NPUs against CPU or GPU execution are what a buyer would want, along with a list of which current laptop and robot applications actually route their AI work to an NPU. This piece cannot offer either.

The paper also could not verify the specifications, availability or capabilities of the new Glenfly chip, the GF800, so it is left out of the picture above.

What to watch

The idea behind the NPU is sound. A specialist beats a generalist on a repeated job, and a battery rewards anyone who does a job with less. Intel putting one in its new Core Series 3 processors, and NXP putting one beside a robot's processor, show the idea is spreading beyond a single product line ^{[3] [4]}.

But the badge is cheap and the software is hard. Watch for tests that show a named feature running on the NPU and report the battery cost. Until they arrive, treat the TOPS number as the start of a question, not the answer.