Last month a friend of mine bought a new Copilot+ laptop, one of those Snapdragon ones everyone was talking about, and within two weeks he messaged me asking why it “feels slower after 20 minutes.” He wasn’t gaming. He was just running Teams with background blur on, editing a video for his team’s internal training, and had about 40 Chrome tabs open like all of us do. Classic case. His shiny new AI PC was throttling.
This is going to be a bigger deal than most people realize. Every laptop and desktop being sold right now is marketed with some AI angle, NPU this, TOPS that, and almost none of the marketing photos mention what happens once the chip actually gets warm. So let’s talk about it properly. Not the marketing version. The version where things get hot, fans get loud, and your expensive new machine quietly gives you less than what you paid for.
I’m not a hardware engineer. I just build my own PCs, mess around with laptops for a living kind of, and have thrown away more thermal paste than I’d like to admit. So this is written from that angle, not from a lab.
What Thermal Throttling Actually Means
Basically, every chip has a temperature ceiling. Intel sets it around 100°C for most CPUs, AMD is a bit more conservative at 95°C. Once a chip hits that limit it begins to throttle, meaning it drops its clock speed automatically to bring the heat back down. It’s not damage. It’s the chip protecting itself. Think of it like your body sweating when you’re overheated, except the chip’s version of sweating is just working slower.
GPUs work a little differently. Most modern GPUs start throttling somewhere around 83 to 87°C, and CPUs typically around 95 to 100°C. The GPU situation is honestly messier because there isn’t one hot spot, there’s the core, the memory, and the power delivery stage (VRM) all generating heat at once. A CPU cooler only has to deal with one central heat source. A GPU cooler has to fight a battle on three fronts.
Here’s the thing that surprised me when I actually looked into this: AI workloads and gaming workloads heat your hardware in completely different ways, even on the exact same chip.
Why AI Workloads Cook Your Hardware Differently Than Games Do
Gaming is bursty. Your GPU spikes to 99% when a big explosion happens on screen, then drops back down while you’re just walking around a map. That constant up and down actually gives your cooler little breathing gaps.
AI workloads don’t do that. If you’re running a local model, doing inference, or even just letting Windows Studio Effects process your webcam feed continuously during a two hour call, the chip sits at a sustained, boring, unrelenting load. No breathing gaps. AI workloads maintain something close to 100% utilization across compute cores for extended periods, creating a slow steady thermal climb instead of the spiky pattern you get from gaming. It’s less dramatic minute to minute but honestly worse for average temperature over an hour.
I didn’t fully believe this until I watched HWiNFO logs myself during a long Stable Diffusion batch versus an hour ofRed Dead Redemption 2. The AI session had a boring, flat, climbing line. The gaming session was jagged, spiky, all over the place. Same laptop. Same room. Totally different heat story.
The NPU Isn’t the Problem, But It’s Not Innocent Either
Now here’s where the “AI PC” branding gets a bit dishonest, or maybe just optimistic. NPUs (the dedicated AI chip that Copilot+ PCs require) are actually really efficient compared to running the same task on your CPU or GPU. AI inference workloads on the NPU run cooler and quieter than most users expect, and ARM based platforms especially tend to produce less heat. That part is genuinely true and I don’t want to be the guy who dismisses it just to sound edgy.
The problem is what happens around the NPU. Microsoft requires 40+ TOPS for the Copilot+ badge, and that number gets tested under ideal lab conditions using a synthetic benchmark that runs in isolation, without competing CPU or GPU activity or memory pressure. Nobody uses their laptop in isolation. You’re running the NPU task while Chrome eats RAM, while Slack notifications fire, while Windows Defender scans something in the background. All of that CPU and GPU activity generates its own heat in a shared thin chassis, and now your “efficient” NPU workload is sharing thermal budget with everything else running at the same time.
So when people say their AI PC gets warm doing “simple stuff,” this is usually why. It’s rarely the NPU alone. It’s the NPU plus everything else fighting for the same small pocket of airflow.
Laptops Have It Worse, and It’s Basic Physics
Desktop PCs have room to breathe. Big case, big heatsinks, multiple fans, actual airflow paths. In a laptop chassis, often less than 20mm thin, thermal management becomes one of the most demanding engineering challenges there is. You’re trying to cram a CPU, GPU, NPU, RAM, and battery into a space the size of a hardcover book and expect it to stay cool during a 30 minute video export.
The trade off is unavoidable. More performance needs more power. More power makes more heat. More heat needs more cooling. More cooling means bigger fans or thicker chassis, which is exactly what laptop buyers say they don’t want. Every thin and light laptop you’ve ever bought made a compromise somewhere in that chain, whether the spec sheet admits it or not.
Business laptops mostly dodge this by using low power chips, 15 to 28 watts, that don’t generate much heat during normal browsing and email. But push them, running a video export, training a small local model, having 60 tabs open while on a video call, and their small cooling systems fall behind fast. I’ve seen a business ultrabook go from silent to sounding like a hair dryer in under two minutes doing exactly this.
Not gonna lie, I used to think fan noise meant a laptop was badly designed. Turns out it’s actually the opposite a lot of the time. A laptop that stays dead silent under heavy AI or gaming load isn’t magic, it’s usually throttling hard and just not telling you. Loud fans that keep clock speeds stable are doing their job. Silent laptops that get warm and sluggish are the ones you should actually worry about.
The RTX 5090 Situation Nobody Fully Explained
If you follow hardware news at all you probably saw the images going around of RTX 5090 power cables measured at over 150°C, hot enough to cause third degree burns if you touched them. A technology journalist in Germany posted thermal images of an RTX 5090 card and its power supply connector showing extreme hot spots where the cable met the GPU. This wasn’t throttling exactly, it was closer to a full blown fire risk from uneven current distribution across the pins in the 12VHPWR connector.
I bring this up because it’s a good reminder that “thermal throttling” and “actual hardware danger” are two different problems that get lumped together in comment sections. Throttling is your chip protecting itself and slowing down. What happened with some 5090 cables was the opposite failure, a connector not designed with enough safety margin for how much power these cards now pull. Nvidia had a similar issue with the RTX 4090 back in 2022, so this isn’t exactly a new lesson, just one that keeps repeating as cards get hungrier for power. If you own a 5090, use the 8 pin to 12VHPWR adapter instead of a single cable pulling everything through one connector, and honestly just check the cable seating every few months. Takes two minutes.
How to Actually Check If You’re Throttling
Most people never look. They just feel their laptop get hot, hear the fans, and shrug. Here’s what I’d actually do instead.
Download HWiNFO, it’s free and it shows real time CPU and GPU temperature alongside clock speed, so you can watch clock speed drop in real time as temperature climbs. On Windows, open Task Manager and check the Performance tab, it now shows NPU usage too if you have one. If you want to check specifically whether Windows is limiting your NPU, open PowerShell and run powercfg /energy, then look for any “NPU Clock Throttling” warning in the report. I didn’t know this command existed until I went digging for this article, so if you didn’t either, you’re not alone there.
For GPUs specifically, don’t just watch the core temperature. Monitoring software usually shows both a GPU core temperature and a junction, or hot spot, temperature, and the junction temperature is always higher and is what actually triggers throttling decisions. I used to only glance at core temp for years and wondered why my card would throttle at readings that “looked fine.” Turns out I was watching the wrong number the whole time.
What Actually Fixes It
There’s no single fix, it depends on what you’re running and what device you have, but here’s roughly what works, ranked by how much effort it takes.
Cleaning dust out of the vents is embarrassingly effective and almost nobody does it often enough. Compressed air every six months, more if you have pets. Second, check your power mode. Most Windows laptops let you switch between Balanced, Best Performance, and a quieter Efficiency mode, and a lot of people leave it on whatever the default was out of the box without ever touching it.
If you’re comfortable going a bit further, undervolting is probably the single best return on effort for anyone doing sustained AI or creative work. You’re telling the chip to run at the same clock speed using less voltage, which means less heat for the same performance. It sounds scary the first time you do it but it’s reversible, and honestly it’s one of the more satisfying things to tinker with once you get the hang of it.
For desktop builders, repaste your CPU cooler every 18 months or so, thermal paste dries out and loses effectiveness over time. And if you’re running a high wattage GPU, check your case has a mesh front panel instead of tempered glass. A full mesh front panel versus a solid glass one can mean a difference of 12 to 15°C in measured GPU temperature, which is honestly a bigger swing than most people expect from something as simple as swapping a case panel.
For laptops there’s less you can do physically, but a cooling pad genuinely helps more than people give it credit for, and keeping the laptop off soft surfaces like beds and couches matters more than almost anything else. I know that’s boring advice. It’s still true.
So Is Buying an “AI PC” Actually Worth It
Here’s my honest take, and I know this is a bit of a hot take itself (no pun intended, okay maybe a little intended). The NPU part of these machines genuinely does what it says, it’s efficient, it runs cool, background blur and live captions and that kind of thing really do sip power compared to doing the same job on the CPU. Where the marketing oversells things is implying the whole machine stays cool and fast just because it has an NPU sticker on it. Your CPU and GPU haven’t gotten smaller or generated less heat just because there’s a new chip sitting next to them.
If you’re buying one of these laptops, don’t shop by the TOPS number alone. Look at thermal reviews for the exact model, not the chip family in general, because two laptops with the identical processor can throttle completely differently depending on how the OEM designed the cooling. That part is on the manufacturer, not the silicon.
My friend ended up fixing his throttling problem by switching his power plan off “Best Battery Life” and cleaning out a surprising amount of dust from a two week old laptop. Turned out the factory had left some kind of protective film stuck over one of the intake vents. Sometimes it really is that dumb of a fix. You don’t always need to undervolt or buy a cooling pad. Sometimes you just need to actually look.