AI 101 — What is an AI PC?

An AI PC is a computer with a dedicated chip for AI work built in — usually a neural processing unit, or NPU — sold on the promise that at least some AI runs on the machine instead of in a data center.

That is the honest definition, and it is narrower than the marketing. The label has been stretched so far since 2024 that it now covers two very different machines: a thin laptop with a small AI chip bolted onto its processor, and a $3,000 desktop-class machine with enough memory to run a 100-billion-parameter model offline. Both get called an AI PC. They are not the same purchase.

Why it matters right now

2026 is the year the category stops being a curiosity. Counterpoint Research expects AI-advanced PC shipments to rise 52% year over year in 2026 and take roughly 59% of global PC shipments, calling 2026 "the crossover year" as NPUs clear the performance bar across mainstream price bands. If you buy a laptop this year, you are probably buying one of these whether you asked for it or not.

The definition most vendors work from is Microsoft's. A Copilot+ PC, per Microsoft's own developer documentation, needs an NPU rated at more than 40 trillion operations per second (TOPS). That number is the whole certification. It says nothing about memory, nothing about which apps use the chip, and nothing about whether the AI features you want run locally or in a server farm.

Meanwhile the top of the market moved somewhere else entirely. Nvidia's DGX Spark mini-desktop is rated around 1,000 TOPS, and the RTX Spark laptops built on the same silicon pair a big GPU with 64 GB or 128 GB of shared memory, PCMag's testing notes. Those machines start near $2,600 and climb past $4,000, How-To Geek found. Same label, twenty times the price, completely different capability.

The mental model: three processors, three jobs

A modern computer already had two brains. The CPU is a handful of very fast generalists — good at anything, great at nothing in particular. The GPU is thousands of slow, simple workers doing the same arithmetic in parallel, which is why it renders graphics and also why it turned out to be good at neural networks.

The NPU is a third kind: fixed-function. It does matrix multiplication — the core operation inside every AI model — in hardware, at low power, continuously, without waking the rest of the chip. That is why it is measured in TOPS and why it is measured in watts. It is built for the always-on stuff: background blur on a video call, live captioning, noise suppression, on-device transcription.

TOPS is a yardstick that flatters. It counts operations, not useful work, and vendors quote the peak number under conditions your spreadsheet will never reproduce. The spec that actually decides what you can run is memory. A model needs roughly its parameter count times its bytes-per-parameter sitting in RAM at once: an 8-billion-parameter model at 4 bits is about 4 to 5 GB, a 30-billion model at 4 bits is closer to 18 GB before you add context. A 40 TOPS NPU with 16 GB of laptop memory cannot run the second one. A slower chip with 64 GB can.

The kitchen analogy

Think of a restaurant kitchen. The CPU is the one skilled chef who can do every station — sauces, grill, pastry — but only one at a time. The GPU is a hundred line cooks: individually unremarkable, collectively able to plate a thousand covers. The NPU is a commercial espresso machine.

That espresso machine is genuinely excellent. It pulls two hundred shots an hour on a fraction of the power, and it never gets tired. It also cannot chop an onion. Buying a kitchen that contains an espresso machine does not make it a restaurant, and buying a laptop that contains an NPU does not make it an AI workstation. It makes it a laptop that is very good at the specific jobs someone wrote NPU code for.

Common misconceptions

"AI PC means the AI runs on your PC." Often it does not. Plenty of features advertised on AI PCs are cloud calls with a local-sounding name, and "available on an AI PC" is not the same claim as "runs on the NPU." Some apps quietly use the GPU instead, because the GPU is faster and its software support is years more mature.

"More TOPS means a more capable machine." Only within the same memory class. TOPS tells you how fast the chip can chew; memory tells you whether the model fits in its mouth. For running a real language model locally, memory capacity and bandwidth matter more than the NPU rating on the sticker.

"You need an NPU to run AI locally." You do not. The mature path is a GPU — a mid-range mobile graphics chip can be rated near 200 TOPS, several times a Copilot+ NPU, and the tooling for it is far better. The NPU's advantage is efficiency, not capability: it does small jobs without draining the battery.

"AI PC is a standard." It is a marketing term with one company's spec attached. Microsoft defines Copilot+; every other vendor uses "AI PC" to mean anything with an NPU in it, from 40 TOPS to 80.

How to judge one

Ignore the badge and ask three questions. How much memory does it have, and is it shared with the GPU? Which of the AI features you actually want run on the device rather than the cloud? And does the software you use target the NPU at all — or will it fall back to the CPU and feel slow? If the answer to the last one is "no idea," you are buying a normal laptop with a spare chip in it, which is fine, as long as you are not paying a premium for the word.

Related reading: AI 101 — What is an NPU (neural processing unit)? explains the chip itself; AI 101 — Local LLMs vs cloud APIs: what's the difference? covers the trade-off the AI PC is trying to win; and AI 101 — What is quantization? explains how models get small enough to fit on one.

Would you pay a premium for on-device AI, or is the cloud good enough? Write to us via our contact page.