You're the person who just got asked whether the new laptops are worth the AI premium, and you don't have a great answer yet. A vendor quote came in with "AI-ready" next to a $200 price bump, and nobody in the room could tell you what that actually buys.
This is part of a series on buying a laptop that can actually run AI locally. If you haven't read the buyer's guide yet, start there: Which Laptops Can Actually Run AI Locally? This post is the one piece of vocabulary underneath that guide worth actually understanding: the difference between a CPU, a GPU, and an NPU.
Three Chips, Three Different Jobs
CPU: does anything, one thing at a time, well
The CPU (central processing unit) is the generalist. It runs your operating system, opens your spreadsheets, handles the logic behind almost everything on the machine. It's built to execute a long sequence of varied instructions quickly, one after another, and it's good at almost any task you hand it. That flexibility is also its limit. Ask a CPU to run the same math operation on a million data points at once and it'll do it, just slowly, because that's not the shape of work it was designed for.
GPU: the same math on thousands of numbers at once
The GPU (graphics processing unit) was built to draw pixels. Rendering a screen means doing the same calculation, over and over, across thousands of pixels simultaneously, so GPUs were designed with hundreds or thousands of small cores that all run in parallel. It turns out that's almost exactly the math AI models need: matrix multiplication, repeated at massive scale. That's why GPUs became the default hardware for training AI models. The tradeoff is power. A GPU pulling that many cores at once draws real electricity, which is fine plugged into a wall and a lot less fine on battery.
NPU: only AI math, but sipping battery
The NPU (neural processing unit) is the newest of the three, and it does one thing: the specific low-precision matrix math that AI inference runs on, and nothing else. It can't run your operating system. It can't render a game. It exists purely to take a trained AI model and run it efficiently, which it does at a fraction of the power draw of asking the CPU or GPU to do the same job. That efficiency is the entire point. It's what lets a laptop run a live transcription, a background blur, or a local AI assistant for hours without murdering the battery.
The Number That Actually Matters: TOPS
NPU performance gets measured in TOPS, trillion operations per second. Microsoft set the bar for its Copilot+ PC designation at 40+ TOPS, and that number is the reason it's worth checking before you buy rather than trusting the sticker. Three current chips clear it in different ways: Qualcomm's Snapdragon X Elite lands around 45 TOPS, Intel's Core Ultra Series 2 (Lunar Lake) is rated around 48 TOPS, and AMD's Ryzen AI 300 series sits at 40+ TOPS. Below that threshold, a machine can still run AI software, just routed through the CPU or GPU instead, slower and hungrier on the battery.
Why This Matters Before You Sign the PO
Here's the part that catches offices off guard: a laptop can have a genuinely great CPU and a genuinely great GPU and still not qualify as an AI PC, because neither of those chips is the one doing the qualifying. The NPU is a separate, distinct piece of silicon, and if it's missing or underpowered, no amount of CPU or GPU horsepower makes up for it. "AI-powered" on a box isn't a spec. The NPU and its TOPS rating are the spec. Everything else is marketing sitting on top of it.
That's the whole distinction, and it's the one number worth asking about before anyone signs off on a fleet refresh.
Not sure which laptops in your quote actually have a real NPU?
We sell Dynabook laptops built around Copilot+ PC-class NPUs. Tell us what your team needs and we'll tell you straight whether the AI spec is real.