Technology · Hardware & devices
AI PCs and NPUs: What's Really Behind the Marketing
“AI PC” is stamped on almost every new laptop. What NPUs actually do, which specs matter when buying – and when waiting for the next generation pays off.
By Boaz Lichtenstein
“AI PC” is the most successful marketing term the PC industry has produced in years – and one of the vaguest. It refers to machines with an NPU, a neural processing unit: a specialised chip that runs AI computations energy-efficiently. The interesting question when buying isn’t whether a device carries the label, but what actually reaches you in everyday use.
Key takeaways
- An NPU isn’t an extra AI feature but an efficiency chip: it handles sustained AI workloads (background blur, transcription) using far less power than a CPU or GPU.
- Cloud AI like chatbots doesn’t benefit from the NPU at all – it only matters for local, continuously running AI features.
- RAM matters more than the NPU’s TOPS figure when buying an AI PC: 32 GB counts as a sensible minimum, because memory is usually soldered in and can’t be upgraded later.
- Microsoft’s “Copilot+ PC” label guarantees a minimum NPU performance level (40 TOPS and up), but not that your software actually uses it.
- Waiting for the next NPU generation is almost never worth it – there’s always a newer one, while fundamentals like RAM and the display determine your satisfaction for years.
What an NPU actually does
An NPU is a specialised chip that computes neural networks with minimal power draw – not faster than a CPU or GPU, but more efficient. CPUs, GPUs and NPUs can all run AI computations; the difference lies in how much battery and heat they burn doing it. The NPU handles sustained-load tasks like background blur on video calls, noise filtering, local transcription or photo indexing without draining the battery or spinning up the fan – tasks a CPU or GPU could also do, just at a multiple of the energy cost.
Performance is measured in TOPS (trillion operations per second) – a figure that markets brilliantly but says little about actual benefit. What matters isn’t theoretical compute power, but which software on your device even talks to the NPU. A chip with a high TOPS number that no installed program uses delivers precisely zero noticeable benefit in everyday life.
The honest cost-benefit case
Most of the AI we use daily – chatbots, research, cloud image generation – doesn’t benefit from an NPU at all, because the computation runs on someone else’s servers regardless. Locally, the NPU scores on three concrete fronts: privacy (transcribing and analysing sensitive content without data leaving the device – a theme closely tied to our article on local AI versus cloud AI), latency (real-time effects without a network round trip), and battery life (continuous AI functions without an energy penalty for the rest of the system).
If none of these three cases applies to you concretely, buying the NPU is mainly buying future-readiness – legitimate, but no reason to pay a noticeable premium over an otherwise identically specced device without a strong NPU. The case only stacks up once at least one of the three benefit points actually shows up in your everyday life.
TOPS and Copilot+: what the number actually means
The TOPS figure on the spec sheet describes raw compute capacity, but says nothing about whether that capacity gets used in everyday life. Microsoft’s “Copilot+ PC” certification currently requires a minimum of 40 TOPS and ties it to certain operating-system features that run locally on the NPU – live captions with translation, or image search by content rather than filename, for example. The label is therefore a useful guide to a device’s baseline specification, but no guarantee that every third-party piece of software actually talks to the NPU.
More important than the absolute TOPS number, then, is the ecosystem check: does the software you use daily – image editing, your meeting tool, operating-system features – actually use the NPU? This question is usually answered within a few minutes of research on the relevant product pages, and it’s more telling than any benchmark bar in a brochure.
What actually matters when buying
- Settle RAM first: 16 GB is the minimum, 32 GB the sensible recommendation – local AI features and models are memory-hungry, and RAM is soldered in on most modern laptops: upgrading it later is impossible.
- Check efficiency before peak performance: the most pleasant devices are the ones with strong performance per watt – noticeable in battery life and fan noise, not in benchmark bars.
- Check your ecosystem concretely: list the three or four AI features you actually use, and check for each one individually whether it runs NPU-accelerated.
- Read the Copilot+ or manufacturer label as a floor, not as the end point of your research – the concrete software list counts for more than the seal.
- Don’t wait for the next generation: there’s always another one. Buy when the need is there – NPU generations evolve quickly, but fundamentals like RAM, display and build quality determine your satisfaction for years. That applies especially if – as described in our article on the right to repair – long usable life matters more to you than the newest label.
- Don’t sacrifice the display, keyboard or battery for the NPU: a device you use all day has to win you over in these categories first – the NPU is the last decision, not the first.
Who actually needs a strong NPU?
Not every type of user gets the same benefit from a powerful NPU. The table below sorts typical use cases by how much NPU performance should actually influence the buying decision:
| User type | NPU relevance | Why |
|---|---|---|
| Heavy video-call users | High | Background blur and noise filters run for hours – battery benefits directly |
| Photo/creative software users | High, if the software supports it | AI filters and object selection run locally, noticeably faster and without a cloud upload |
| Gamers | Low | Frame rate and graphics depend on the GPU, not the NPU |
| Standard office work | Low | Chatbots and cloud AI run server-side, independent of the local NPU |
| Developers running local models | Medium | RAM and GPU memory are often more limiting than the NPU alone |
From experience: the most reliable test before buying is an in-store check or a ten-minute look at a demo unit: start a video call with background blur and watch the battery indicator. If it drops noticeably faster than on a comparable device without active NPU use, the software isn’t NPU-optimised – regardless of what the spec sheet says. If you don’t have a demo unit to hand, you’ll often find the same information in current tech-press reviews that measure exactly this battery drain under AI load.
The most common mistakes when buying an AI PC
- Basing the buying decision on the TOPS number alone. Fix: first check whether your own software even talks to the NPU – the number is meaningless without software that uses it.
- Skimping on RAM because “16 GB is surely enough”. Fix: choose 32 GB if local AI features matter to you – upgrading later isn’t possible on most models.
- Waiting for the “next, better generation”. Fix: buy based on your current need; the cycle of new NPU generations is short enough that waiting is rarely rewarded.
- Confusing NPU performance with gaming performance. Fix: the GPU still decides for games – a strong NPU doesn’t turn a device into a gaming machine.
- Treating a certification label as the whole buying decision. Fix: read the label as a minimum standard, then still check the concrete software list.
The bottom line
The AI PC isn’t a revolution but a solid evolution: more efficiency for tasks that, in some cases, have been around for a while already. The NPU pays off concretely if you regularly use video calls, AI-powered creative software or local transcription – and is purely a future option if you don’t. Anyone who settles RAM, the display and battery first, then answers the NPU question based on their own software, automatically makes the right call – regardless of the label on the box.