NPUs Are in Over Half of New PCs. Almost Nothing Uses Them

AI-capable PCs now account for more than half of global laptop shipments, according to Gartner — yet on most of those machines, the NPU at the heart of the certification sits idle. On a typical Windows 11 machine, Task Manager shows its utilization at zero percent.

Why? The hardware isn't there because buyers demanded it. It's there because Microsoft dictated it. In 2024, the company told laptop makers that new machines need 40 TOPS of dedicated NPU performance — TOPS being the measure of how fast an NPU chews through AI math — or they lose access to new Windows AI features. So the silicon shipped. The software didn't follow.

What does an NPU actually do?

An NPU, or Neural Processing Unit, accelerates AI inference — the "run this model" phase, not the GPU-driven "train this model" phase. It's purpose-built for the two operations neural networks lean on: matrix multiplication and nonlinear activation functions. That specialization buys speed and efficiency at the cost of the general-purpose flexibility a CPU or GPU offers.

The power math is the selling point. An NPU runs AI workloads at roughly 2W, versus well over 100W for a discrete laptop GPU doing the same job. Think background blur on a video call: it can run on a CPU or GPU, but both generate heat and sap battery life. On an NPU, the same effect runs continuously without spinning up the fans.

Don't expect miracles, though. NPUs can't train models, and running large local LLMs like Llama 70B still demands a discrete GPU with serious video memory. ChatGPT and Gemini don't need your NPU either — those run on provider infrastructure. The real use cases today are narrow: webcam effects, noise suppression, real-time translation, and local search indexing.

How did we get a 40 TOPS floor?

Microsoft's Copilot+ PC certification sets the bar: 40 TOPS from the NPU specifically, plus 16GB of RAM and 256GB of storage. Note the fine print — the 40 TOPS must come from the NPU, not the GPU. A mid-range laptop RTX 4050 delivers north of 190 TOPS from its tensor cores, and it doesn't count.

The spec triggered a silicon race:

• Qualcomm's Snapdragon X Elite and Plus shipped first with a 45 TOPS Hexagon NPU
• Intel's Lunar Lake followed with up to 48 TOPS
• AMD's Ryzen AI 300 hit 50 TOPS
• Intel's Panther Lake (Core Ultra Series 3) lands at 50 NPU TOPS on the 18A process
• Later this year, Qualcomm's Snapdragon X2 Elite reaches as much as 85 TOPS — more than double Microsoft's minimum

Every generation ratchets the number higher even though software to consume those TOPS remains scarce. Dell has already stuffed an enterprise-grade NPU into its Pro Max Plus. Soon, AI laptops may be the only option on the shelf, which makes the NPU unavoidable whether you use it or not.

The app list that actually leans on the NPU is short: Windows Studio Effects, Live Captions, Recall, Click to Do, Teams and Zoom noise suppression, and a handful of Adobe Creative Cloud features.

What does an NPU look like under load?

One hands-on test shows both the promise and the friction. On a 2024 ASUS Zenbook Duo UX8406 with 32GB of RAM, running the open-source Mistral 7B model through Intel's OpenVINO toolkit pegged the Intel AI Boost NPU at 98 percent utilization while the CPU sat at three percent. Text generation worked, the machine stayed responsive — this on a platform with 34 combined TOPS, only 11 of which are dedicated to the NPU.

Run the same model via Ollama on the CPU instead, and utilization jumps to 52 percent with memory use around 17GB of the 32GB capacity. Add any other CPU-heavy workload and the laptop would grind to a halt. That's the NPU's real pitch: it handles AI work without touching the rest of the system.

The catch? Getting there required a specific toolkit and technical know-how. Native NPU support in consumer software is still close to nonexistent — and until that changes, half the world's new laptops are carrying silicon most users will never wake up.