RTX Spark: will your next Windows PC run AI without the Internet?

“Without Windows, there would be no GeForce.” Jensen Huang, Nvidia's boss, said that on Wednesday October 7 in San Francisco, sitting next to Satya Nadella, Microsoft's boss. Nice compliment, and not a free one: the two had come to present RTX Spark, a new chip for Windows laptops. The stated goal: run an AI directly on your computer, without the Internet, without a subscription, and without your files being sent to someone else. On paper, it looks a lot like what I've been waiting for for months. But let's take a closer look.

Router unplugged, and the AI still replies

One big pantry instead of two small fridges

RTX Spark is the processor and the graphics card soldered onto the same chip, with up to 128 GB of memory shared between the two. And it's this shared memory that changes everything.

In a classic PC, the processor has its memory, the graphics card has its own, and they don't lend each other anything. Two small fridges in two different rooms. But for an AI to work, it has to fit entirely in the graphics card's memory, and a good gaming card has 16 GB. That's too small for big models. With RTX Spark, everyone digs into the same big 128 GB pantry. Apple has been doing exactly that on its Macs since 2020, and that's actually why so many AI tinkerers have a MacBook.

The difference between separate CPU and GPU that share nothing, and one single pantry

And during the presentation, Nvidia ran Qwen 3.8 Flash Next, an AI model with 125 billion parameters (the small settings an AI learns during its training), directly on the machine, offline. To give you an idea, in August I was telling you about a Meta model that fits on a graphics card, and that was already quite a little party. Here, we're talking about a model several times bigger, in a laptop.

The agent sleeps in the guest room

The least spectacular part may be the most important one. Microsoft released a system called MXC that same day, for Microsoft Execution Containers. An “agent” is an AI that doesn't just answer: it acts, it opens files, it fills in forms, it launches programs. Practical. And a little scary when you let it have the keys to the whole house.

MXC gives it a guest room. In the strictest mode, the agent runs under its own Windows account, with its own desktop and its own clipboard, and it only sees what its rules allow. The rules are written outside the agent, so it can't give itself more rights on its own. And when it does something stupid, the history clearly says it was him and not you. OpenAI's Codex and GitHub Copilot already handle this, Claude Code should follow.

One key only, the agent stays in its place

This lines up with what Apple is doing on its side, which is going to lock down full access to the disk because of AI agents. Apple and Microsoft have therefore reached the same conclusion in the same week: you don't give an agent access to the whole computer. And for that, I say bravo! Microsoft even specifies that MXC works on existing PCs, not just on new machines.

Concretely, what does this change for you?

Imagine your tax return, your pay slips, the house papers, the kids' photos. Today, if you want an AI to help you sort through all that, you send it to a company's servers, somewhere. With an AI running on your PC, nothing leaves the office. You can even unplug the internet box, it still works. On the train, in a tunnel, in the chalet with no network.

And then there's the money. Me, watching my Claude Code quota every five hours like a hawk, I can immediately see the appeal of an AI that doesn't count anything: you've paid for the machine, and you can use it as much as you want.

She gets on with her paperwork. He's still looking for a signal bar

But not right away, and not for everyone. The starting price is $2,599 in the United States for Microsoft's Surface Laptop Ultra, and it climbs quickly when you add more memory. There isn't an official price in euros yet, but with European taxes, it won't be cheaper. The first machines come out on October 16, from Microsoft, Acer, ASUS, Dell, HP, Lenovo, MSI and Gigabyte.

The good news is, we've seen this before. When Apple released its first in-house chips with shared memory, in 2020, it was limited to a few models. Today, it's in every Mac, including the cheapest ones. And 3D cards in the 1990s, reserved for wealthy gamers, are now in any phone. This kind of tech always works its way down the price range. The question is when: two years, four years, nobody can promise you.

The small print

First point: the chip is built on the ARM architecture, the one used by phones and Macs, and not on the one used by regular Windows PCs. Most old Windows software goes through a translator, called Prism, which makes it run anyway, but a little more slowly. For online games, that was the big sticking point: anti-cheat software refused to start. Nvidia says it has sorted that out with the publishers, and Riot announces that League of Legends and Valorant will work. I want to see the first tests.

Second point: all the speed figures come from Nvidia itself, including those comparing it with the MacBook Pro. The real independent tests will arrive with the machines, on October 16. And even the Engadget site, which got its hands on Microsoft's laptop and finds it superb, points out that it isn't made for the general public but for AI developers.

My opinion

For me, this is the future. For nearly four years now, AI has lived in the cloud, with three or four companies setting the prices, quotas and rules while we pay the rent every month, and then suddenly we're told that AI can come back to your desk, into your machine, with a set of rules stopping it from touching everything, and that it will still work the day your internet box breaks down. I'm in! On the other hand, at $2,599 for the entry ticket, I'll wait for the second generation, as usual, and I advise you to do the same.

And to Apple, which had shared memory and the silent laptop all to itself for five years: you've got company! Maybe it's time to lower the price of memory on your MacBooks!

Sources

• Nvidia, October 7, 2026: RTX Spark and AI agents on Windows
• Microsoft, October 7, 2026: Microsoft Execution Containers, the rules governing agents
• Engadget: hands-on with the Surface Laptop Ultra (price, intended audience)
• Tom's Hardware: RTX Spark, ARM and anti-cheat software for games

Article written with the help of Claude Code, proofread and corrected by me.