Microsoft announced Copilot+ at an event in Redmond today. The new brand is for AI PCs, with Microsoft reimagining how people interact with PCs in an era with emerging AI technologies.

It's all about bringing AI from the cloud to your device, running locally on an neural processing unit, or NPU. Of course, local AI tasks aren't new, it's just that most PCs have been using GPUs to tackle them.

A dedicated GPU won't get you local Copilot+

There are three requirements to get Copilot+:

• An NPU capable of 45 TOPS or more
• 16GB RAM
• 256GB storage

If you go and look up the specs for an Nvidia GeForce RTX 4090, you'll find that it's capable of 1,321 "AI TOPS", which I put in quotes because not all TOPS are equal and the specs page doesn't say if Nvidia is referring to INT8 or something else, although with a number that high, it really doesn't matter. Qualcomm and Microsoft are using INT8, for reference.

The point is, an RTX 4090 runs circles around a 45TOPS NPU, no matter how you slice it. However, it won't be enough to get you Copilot+. Those specs specifically refer to an NPU, and Microsoft is committed to that, at least for now.

So, what gives?

I asked quite a few people at Microsoft about this, as I've spent the week attending its event in Redmond, followed by Build in Seattle. After all, I've been using my own RTX 4080 for all kinds of AI tasks like testing out LLMs, and after I started reviewing Intel Meteor Lake laptops, which only come in at around 11 TOPS, it was clear that a desktop GPU is going to outgun laptops for a long time to come.

The best answer I got was that Microsoft is just focusing on laptops right now. If you've got an RTX 4090, or even a GPU from a generation or two behind that, it's powerful enough, but you're an outlier. Laptops are mainstream, so the Redmond firm targeted the biggest chunk of its base, by far.

It's also not as simple as flipping a switch. These features are coded to target NPUs, so if you just lit up the feature on other PCs, they're really not optimized for that.

This isn't to say it's not coming. And no, Microsoft isn't going to publicly comment on upcoming features, so there are no promises to be made here. I was assured by some high-ranking folks on the Windows team that Microsoft has no intention of leaving desktop users behind, and it's absolutely thinking about them.

But if Microsoft doesn't ship its on-device AI features from Copilot+ to run on dedicated graphics, there are other alternatives coming later. Intel's next-gen 'Arrow Lake' desktop CPUs are going to come with on-board NPUs, so if nothing else, that would being Copilot+ to the desktop.

Also, graphics cards are going to start shipping with on-board NPUs as well. I'm not sure how that's different from the Tensor units found in existing Nvidia GPUs, so we'll have to wait and find out on that one.