AI Laptop Features Explained: What NPUs, TOPS, and Copilot+ Actually Do
By Chester Takau · July 2026
TL;DR
- An "AI laptop" has a dedicated NPU (neural processing unit) chip for running AI tasks locally instead of in the cloud
- TOPS (trillion operations per second) measures NPU speed — Microsoft's Copilot+ PC standard requires 40+ TOPS
- Real features today: on-device transcription, background blur/eye contact in webcams, local image generation, offline voice typing
- Windows Recall and similar "remember everything" tools are the most-marketed but least battle-tested feature
- Most people do not need to buy a laptop for the NPU alone yet — check what runs on CPU/GPU vs what needs the NPU before paying the premium
"AI laptop" shows up on nearly every product page in 2026, but the term covers two very different things: a chip that runs AI tasks locally (the NPU), and marketing labels attached to features that mostly run in the cloud anyway. This guide breaks down what the NPU actually does, what TOPS means, which features genuinely need the chip, and which are running the same cloud AI your laptop had two years ago with a new sticker on the box.
What makes a laptop an "AI laptop"
The defining hardware is the NPU — neural processing unit. It is a chip built specifically to run the matrix-multiplication math that AI models use, separate from the CPU (general processing) and GPU (graphics and parallel compute). NPUs exist because that specific type of math can be done far more efficiently — using a fraction of the power — on dedicated silicon than on a CPU. A laptop needs a physical NPU on the motherboard to qualify; a laptop that just runs ChatGPT in a browser is not an "AI laptop" by this definition, even if the marketing says otherwise.
TOPS explained — the number on every spec sheet
TOPS stands for trillion operations per second, and it is the standard measure of NPU throughput. Higher TOPS means the chip can run larger AI models, or the same model faster, using less battery. Microsoft set 40 TOPS as the minimum bar for its "Copilot+ PC" certification, introduced in 2024 — a laptop with less NPU power than that cannot carry the label or run the Copilot+ exclusive features, regardless of how fast its CPU or GPU is.
NPU performance by platform (2026 laptop chips)
| Chip family | NPU TOPS | Copilot+ certified |
|---|---|---|
| Qualcomm Snapdragon X Elite / X2 Elite | 45–80 | Yes |
| Intel Core Ultra 200V (Lunar Lake) | 48 | Yes |
| AMD Ryzen AI 300 / 400 series | 50–55 | Yes |
| Apple M4 / M5 | ~38 (16-core Neural Engine) | N/A — uses Apple Intelligence instead |
| Intel Core Ultra 100 series (Meteor Lake, 2024) | 10–11 | No |
A high TOPS number is only useful if software is written to actually call the NPU. Right now that gap is the real bottleneck — the hardware in most 2025–2026 laptops outpaces the number of applications built to use it.
Features that genuinely use the NPU
These run locally, work offline, and measurably benefit from dedicated AI hardware:
- Live Captions with translation (Windows) — transcribes and translates audio from any app, entirely on-device, with no internet connection required
- Windows Studio Effects / webcam AI — background blur, eye contact correction, and auto-framing run continuously during video calls without spiking CPU usage or draining the battery the way software-only versions do
- Voice typing and dictation — local speech-to-text models (similar in concept to Whisper) that work without sending audio anywhere
- Cocreator and local image generation — small on-device image models for quick sketches-to-image inside Paint and Photos, at lower quality than cloud tools but instant and offline
- Apple Intelligence writing tools on M-series MacBooks — proofreading, summarizing, and rewriting text run on the Neural Engine without a network round trip
Features marketed as "AI" that mostly are not NPU-dependent
Windows Recall — the feature that screenshots your activity and lets you search it later in plain language — does use the NPU for on-device indexing, but it is the most scrutinized AI PC feature so far. It shipped later than planned after a security backlash over unencrypted screenshot storage, and the shipped version requires Windows Hello enrollment and encrypts the database at rest. It is real NPU usage, but it is also the feature with the most legitimate privacy questions attached, and it is opt-in rather than on by default. Most chatbot integrations baked into a laptop's default apps — the built-in Copilot sidebar, browser AI assistants, most "AI search" features — call a cloud model over the internet. The NPU has nothing to do with those; a laptop from 2019 running the same browser gets the same result, just slightly slower to render.
"If a feature works with the WiFi off, it is using the NPU. If it stops working, it was a cloud feature wearing an AI PC sticker."
Why on-device AI matters beyond speed
The three practical advantages of NPU-based features over cloud AI are consistent across platforms: they work without internet, they do not send your audio, screenshots, or documents to a server, and they cost nothing per use since there is no API call being billed. The tradeoff is capability — a local model small enough to run on a laptop chip is meaningfully less capable than a large cloud model like GPT-5 or Claude. On-device AI is built for narrow, fast, private tasks (transcribe this, blur this background, caption this call), not for open-ended reasoning or long-form generation. Laptops increasingly combine both: local NPU for quick, private, always-on tasks, and a cloud subscription for anything that needs a frontier-scale model.
RAM matters more than most buyers realize
Any on-device AI model — even a small one — has to load into memory to run. Microsoft's Copilot+ PC certification requires a minimum of 16GB RAM specifically because local AI features share that pool with everything else open on the machine. Buying an "AI laptop" with 8GB of RAM defeats the purpose; the NPU will be starved of memory the moment more than a browser and one AI feature are running. If you are buying specifically for the AI features, 16GB is the practical floor and 32GB is worth the extra cost if the laptop will also run local large language models for coding or writing assistance.
Do you actually need to buy one
If your current laptop already runs Claude, ChatGPT, or Gemini fine in a browser, an NPU adds nothing to that experience — those run in the cloud regardless of local hardware. The NPU earns its keep for three groups: people on video calls all day who want background effects without battery drain, people who need captions or transcription without uploading audio anywhere, and people who want a faster, quieter laptop overall since NPU offload reduces fan noise and heat during everyday multitasking. For everyone else, buying based on CPU performance, screen quality, and battery life — with the NPU as a bonus rather than the deciding factor — is still the better use of a laptop budget in 2026.
For the underlying concepts behind these features, what is machine learning explains how the models themselves work, and what is computer vision covers the tech behind webcam effects like background blur and auto-framing. If you are deciding between AI hardware categories, the best AI wearables 2026 guide covers the wrist- and face-worn side of the same trend, and tech that survives the Pacific Islands covers what actually holds up in daily use once you own the hardware. For the software side of AI once you have picked a laptop, best AI tools for beginners is the next stop.