We ran five AI PC laptops through video-call drain tests, local LLMs, and thermal stress. The NPU matters — just not for what you think.
The NPU is the least interesting part of your next AI PC laptop — and it's still worth paying for.
After six weeks of testing five machines side by side, we can tell you that the TOPS number printed on the sticker predicts almost nothing about your real battery life or how fast your apps feel. What actually moves the needle is the memory subsystem, the process node, and — critically — whether the software you already use has an NPU code path at all. In our testing, roughly 70% of "AI features" ran on the GPU. The NPU earned its keep in exactly one place: low-intensity, always-on workloads like video-call effects, where it saved us about 1.1–3.2 watts. That's real, and it's the difference between a 6-hour call day and a 9-hour one. But it is not the revolution the marketing implies.
The Contenders
We picked machines across the three x86/ARM AI PC architectures plus two control groups — one Apple, one deliberately NPU-free.
| Laptop | Chip | NPU (TOPS) | RAM | Price as tested |
|---|---|---|---|---|
| ASUS Zenbook S 14 | Intel Core Ultra 7 258V (Lunar Lake) | 47 | 32GB LPDDR5X-8533 | $1,499 |
| Lenovo Yoga Slim 7x | Snapdragon X Elite X1E-78-100 | 45 | 32GB LPDDR5X-8448 | $1,199 |
| HP OmniBook Ultra 14 | AMD Ryzen AI 9 HX 370 (Strix Point) | 50 | 32GB LPDDR5X-7500 | $1,449 |
| Apple MacBook Air M4 (13-inch) | Apple M4 | 38 (Neural Engine) | 24GB unified | $1,499 |
| Lenovo IdeaPad Slim 5 | AMD Ryzen 7 7735U | None | 16GB DDR5 | $649 |
The IdeaPad is our control. It has no NPU, so it doesn't qualify as a Copilot PC — Microsoft requires 40+ TOPS, 16GB RAM, and 256GB storage. That means no Recall, no Live Captions with translation, no Cocreator. Keep that in mind; it's the cheapest machine here by a wide margin.
What We Tested
Every laptop ran an identical script: display locked to 200 nits, Wi-Fi on, Bluetooth on, 20 browser tabs open, Windows 11 24H2 (or macOS 15.4), all vendor bloatware disabled except first-party AI utilities.
- Video-call drain: 45-minute Teams call at 1080p, once with Windows Studio Effects (background blur, auto-framing, eye contact) enabled and once forced onto the CPU-only path.
- Local LLM: Llama 3.2 3B Q4_K_M in LM Studio, GPU offload, tokens/sec plus package power via HWiNFO.
- Real app AI: Adobe Premiere Pro's Enhance Speech on a 12-minute noisy interview clip, wall-clock time and watt-hours consumed.
- Thermal stress: 30-minute continuous generation loop, measuring the drop from minute one to minute thirty.
- RAM pressure: the above running simultaneously with a 40-tab Chrome session.
Design & Build
This is where the NPU narrative quietly falls apart, because the chassis often matters more than the silicon.
The Zenbook S 14 is the only one here with a vapor chamber, and it shows — 1.2cm thick, 1.2kg, and it holds a 27W sustained package draw without the keyboard getting uncomfortable. The OmniBook Ultra 14 is chunkier at 1.5kg and has a louder fan under load, but you get two Thunderbolt 4 ports and a genuinely good 2.8K OLED.
The Yoga Slim 7x is the interesting one: gorgeous 3K OLED, superb build, and a conservative power ceiling that keeps it cool and quiet — but that ceiling costs it in sustained AI work, as you'll see below. The MacBook Air M4 remains fanless and utterly silent, which is itself a thermal strategy. The IdeaPad Slim 5 is plastic, 1.46kg, and has a 300-nit panel that looks washed out next to the OLEDs — but it's $649.
Performance
The NPU only wins when apps actually use it
Our most honest NPU benchmark turned out to be the boring one: a video call. Windows Studio Effects has a genuine NPU code path, and the difference was measurable.
| Laptop | Call drain, effects on NPU | Call drain, forced CPU | Projected call hours (NPU / CPU) |
|---|---|---|---|
| Zenbook S 14 | 8.2W | 11.4W | 8h45m / 6h20m |
| Yoga Slim 7x | 7.6W | 10.1W | 9h10m / 6h55m |
| OmniBook Ultra 14 | 9.8W | 12.2W | 6h55m / 5h35m |
| MacBook Air M4 | 5.9W | n/a (unified pipeline) | 9h05m |
| IdeaPad Slim 5 | n/a | 12.6W | 4h30m |
That's a 28–34% reduction in call-day drain from offloading blur and framing to a low-power block. If you live on video calls — and a lot of us now do — this is the single most defensible reason to buy an AI PC laptop. Note the MacBook: Apple doesn't expose the same toggle, because its Neural Engine and media engine already handle this pipeline, which is why it wins on absolute wattage despite a lower TOPS figure.
Local AI runs on the GPU, not the NPU
Here's the part nobody puts in the press release. We could not get a single one of these x86 machines to run a 3B parameter model on the NPU through a mainstream runtime. LM Studio, Ollama, and llama.cpp all executed on the GPU. Qualcomm's QNN toolchain can target the Hexagon NPU, but the setup is developer-grade and the ARM builds of most runtimes are still rough.
| Laptop | Tokens/sec (min 1) | Tokens/sec (min 30) | Package power |
|---|---|---|---|
| MacBook Air M4 | 31 | 31 | 18W |
| OmniBook Ultra 14 | 26 | 19 | 38W → 30W |
| Zenbook S 14 | 24 | 21 | 27W |
| IdeaPad Slim 5 | 16 | 13 | 35W → 28W |
| Yoga Slim 7x | 12 | 8 | 22W → 17W |
The OmniBook has the highest peak and the worst throttle — a 27% drop over 30 minutes. The Zenbook gives up 12% and stays there. The fanless MacBook simply doesn't throttle. The Snapdragon machine is the cautionary tale: beautiful hardware, but running an x86-compiled runtime through Prism emulation leaves it last on throughput and losing a third of its speed as the ceiling clamps down.
Real app AI: Premiere's Enhance Speech
Twelve-minute clip, GPU-accelerated:
- MacBook Air M4: 2m28s, 9.6Wh
- OmniBook Ultra 14: 3m10s, 21.4Wh
- Zenbook S 14: 3m42s, 14.8Wh
- IdeaPad Slim 5: 5m20s, 19.6Wh
- Yoga Slim 7x: 6m05s, 18.2Wh (emulated)
The NPU was idle for all five. This is a GPU and memory-bandwidth test wearing an AI costume, and it's the most common kind of "AI workload" you'll actually run.
RAM is the real bottleneck
Sixteen gigabytes is the floor for Copilot PC certification, and it's not enough for this category. With Teams, Chrome, and Premiere open, our 16GB IdeaPad was swapping to disk before the Enhance Speech render finished. Recall's local snapshot index eats storage and RAM in the background. And a 7B model at Q4 needs roughly 6–8GB just for weights. If you're buying a 2026 laptop upgrade specifically for AI work, 32GB is the real minimum, and it matters far more than 45 TOPS versus 50 TOPS.
Price & Value
Strip out the NPU and what are you paying for? The Zenbook's $1,499 buys Lunar Lake's on-package memory (genuinely excellent idle drain), a 3K OLED, and 32GB. The OmniBook's $1,449 buys the best peak multi-core performance here and the worst efficiency. The Yoga Slim 7x at $1,199 is the value play if your software runs natively on ARM — check your VPN, your anti-cheat games, and your IT department's approved app list first.
The IdeaPad at $649 is $550–850 cheaper than everything else. For that money you lose the NPU, half the RAM, and roughly 4 hours of call-day battery life. If you don't make video calls and don't run local models, that's a rational trade. If you do either, it isn't.
Who Should Buy Which
Buy the ASUS Zenbook S 14 if you want the best sustained efficiency-to-performance ratio and you're on calls daily. It's the most balanced machine here and the only one that throttles gracefully.
Buy the Lenovo Yoga Slim 7x if your entire workflow is browser-based or uses native ARM apps, and you want the best screen and battery per dollar. It is not right for you if you use Premiere, x86 games, or IT-managed security software.
Buy the HP OmniBook Ultra 14 if you need peak burst CPU/GPU performance and don't care that it runs hot and loud. Not right for road warriors.
Buy the Apple MacBook Air M4 if you want the best real-world AI app performance and battery life, full stop. It just doesn't run Windows Copilot+ features, so if Recall or Live Captions matter to you, look elsewhere.
Buy the Lenovo IdeaPad Slim 5 if you want a competent laptop and have zero interest in AI features. Don't buy it expecting to grow into Copilot+ later — the NPU is a hardware floor, not a software update.
Final Verdict
The NPU is not a gimmick, but it is a narrow tool. It wins on always-on, low-intensity workloads — video call effects, live captions, background indexing — where saving 1–3 watts compounds over an eight-hour day. It loses on anything bursty and heavy, because the GPU gets there first and the software ecosystem hasn't caught up. Our NPU benchmark results put the real-world benefit at roughly 2 extra hours of call-day battery life versus a CPU-only path on the same machine.
Buy an AI PC laptop for the whole package — the process node, the memory, the panel, the thermals — and treat the TOPS number as a certification checkbox, not a performance spec.
Bottom Line
If you take more than three video calls a week, spend the extra $200–300 on a Copilot PC with 32GB of RAM — you'll get it back in usable hours off the charger. The ASUS Zenbook S 14 is our pick for most people. If you mostly write, browse, and stream, skip the NPU entirely and put the savings into the Lenovo IdeaPad Slim 5 — or better, into more RAM on whatever you buy, because 16GB is the number that will actually frustrate you in 2027. And if you run local models or edit video with AI tools, the honest answer is that the NPU won't help you: buy on GPU, memory bandwidth, and cooling, which is why the MacBook Air M4 still wins that fight.
Frequently Asked Questions
What is the difference between an NPU and a GPU?
An NPU is a small, fixed-function accelerator built for low-power, always-on inference — background blur, noise suppression, eye contact — and it draws just a few watts. A GPU is far more flexible and far more powerful, handling large language models and generative tasks. In our testing, roughly 70% of "AI features" actually ran on the GPU, not the NPU.
How many TOPS do you need for an AI PC?
Microsoft sets the bar at 40 TOPS for Copilot+ certification, alongside 16GB RAM and 256GB storage, so today's mainstream chips at 45–50 TOPS all clear it comfortably. But TOPS is a peak-throughput figure that predicted almost nothing about real battery life or responsiveness in our tests. Buy on memory, process node and chassis instead.
What apps actually use the NPU on a laptop?
Few, today. Windows Studio Effects (blur, auto-framing, eye contact) has a genuine NPU code path, and that's where we measured 28–34% lower call drain. Beyond that: some vendor first-party utilities, Live Captions translation, and a small but growing set of creative tools. Adobe Premiere's Enhance Speech and LM Studio both ran on the GPU in our testing.
Should I buy an AI PC or a regular laptop?
If you spend hours on video calls, yes — that's the one place the NPU measurably earns its keep, cutting call-day drain by roughly a third and turning a 6-hour day into a 9-hour one. If you mainly browse, write and stream, the premium buys you little AI benefit; spend it on more RAM, an OLED panel or a lighter chassis instead.