Microsoft’s new Surface Laptop Ultra has a presence. Physically, it’s a gorgeous device with a 2,000-nit, 15-inch mini-LED display, a large haptic touchpad, and a stalwart build that looks both powerful and expensive (both are true). If ever there were a device made for power users, this is it.

But there’s also a specific power user in the crosshairs here. Microsoft is hoping to lure the MacBook Pro user who wants not only something premium but a device capable of running the next generation of local AI. With Nvidia’s new 20-core RTX Spark N1X chip and up to 128GB of unified storage, it’s built for the future, indeed.

The sheer power is the draw here, since the Surface Laptop Ultra is not more affordable than a MacBook Pro, nor does it offer better CPU performance for everyday tasks. It’s a Windows machine, so you’ll have more gaming options, but it’s also running on Arm, which has its own concerns. Essentially, the ideal Surface Ultra user is so committed to this laptop’s advertised use case as a local AI beast that they see no other option.

Impressively ‘Ultra’

The Surface Laptop Ultra has an exceptional build, and while not exactly lightweight at 4.4 pounds, it still feels well-balanced and durable, with a large haptic touchpad and rounded edges that resemble the MacBook Pro in several ways. The 15-inch mini-LED touchscreen features a 3:2 aspect ratio, a 120Hz refresh rate, and up to 2,000 nits of brightness — the brightest Surface laptop to date. It’s gorgeous and exceptionally glossy, but also bright enough to override glare.

It has an ample selection of ports: three USB-C ports, a USB-A, an SD card reader, a headphone jack, and an HDMI, with the right-side USB-C featuring the viral magnetic snap-to feature that Microsoft kept secret until its October Surface event in San Francisco. It’s not a game-changing feature, but it is a welcome quality-of-life touch.

Speaking of quality of life, it also supports Microsoft’s new haptic signals in Windows, which trigger on basic UI actions like resizing windows and aligning objects, and are available in a few limited apps (for now). Microsoft hasn’t confirmed specifics, but it’s likely we’ll see additional apps and devices support it soon, given that these effects live at the OS level.

Other features on the Surface Laptop feel mixed. The 1080p Surface Studio Camera looks fantastic; it’s one of the best webcams I’ve used on a PC this year. The keyboard and trackpad are superbly ergonomic, and it’s a laptop you can work on for hours at a time. However, it takes a long time to boot up and restart, and I had issues with the USB-C ports generating random dialog boxes in Windows every few moments, even when it was just the charger.

Also: I put Lenovo’s RTX Spark laptop through a bonkers AI coding test

The battery life is also surprisingly good, easily lasting longer than a workday on one charge, even while running demanding tasks. Battery performance is one of those features we shouldn’t overlook, even though we’re spoiled for such good battery endurance in 2026. As I used this laptop for an entire week, I barely worried about battery life.

The future of Windows (?)

Nvidia says its new N1X chip in the RTX Spark laptops is a “new beginning for Windows PCs,” built for creators, gamers, and most importantly — AI agents. The model I reviewed powered 64GB of unified memory and a 1TB SSD, putting the starting price at $4,299 (or $5,899 for the 128GB config).

Nvidia’s chip runs Windows on Arm architecture, with a 6144-core Blackwell RTX GPU and a 20-core CPU. It scales up to 128GB of unified LPDDR5x memory, with as little as 24GB on the low end. Storage goes up to 2TB.

Also: Touchscreen MacBook Pro hands-on: A stunning OLED

That’s a whole lot of power, but this is no general-purpose laptop. It’s for developers and programmers, creators, and AI power users on the frontier looking to harness agentic AI locally with powerful hardware tailored to those use cases. The thing is, though, all that powerful hardware doesn’t necessarily translate to daily workflows in ways you might think.

As mentioned, I took the Surface Laptop to the office for a week and didn’t find its performance for basic multitasking to be particularly mind-blowing. Multiple apps open at once with several dozen tabs often resulted in inconsistent performance, especially when connected to multiple displays. I had glitches with the USB-C port popping up a ghost dialog box every few moments, only to disappear. After updates and switching to best performance mode, I noticed a difference, but it raises the question of how optimized the hardware is.

Kyle Kucharski/ZDNET

When it comes to single-core performance, the MacBook Pro M5 and Mac Mini M5 Pro significantly outpace it in our benchmarking tests, as evidenced by certain tasks.

When running local AI, however, this thing is made for it, right from the first boot-up. After setting up Windows, you’re prompted to install an optional local agent: Hermes, OpenClaw, or Perplexity. All of them offer customizable options for configuring local AI on your machine and an accessible way to get started. That said, getting everything off the ground running wasn’t exactly seamless.

Running agentic AI locally

I have many years of experience working with Windows, but I’m still learning how to adopt local agents into my daily workflows. As someone who’s not a developer, I found that getting OpenClaw to work out of the box required some trial and error. First, I had trouble getting the gateway to successfully connect, only to find that I needed to install WSL (Windows Subsystem for Linux). This was required because the OpenClaw local Gateway setup uses a Linux environment managed through WSL, but for whatever reason, it wasn’t in the version of Windows I was running.

Also: I tried local AI on HP’s Omnibook Ultra – it’s a private, powerful beast

After over an hour, it was all set up with the llama.cpp local inference engine, I downloaded the Qwen3.8-27B model, and we were off. Since this was a loaner device, I went the YOLO route and gave the agent access to everything on my PC (but don’t recommend doing that in practice). I had a specific task in mind: organize a folder of 100 or so loose .jpg images into new folders based on their content.

It proved to be a more complex task than I thought, which is part of why there’s a fair amount of skepticism for these models. But in short, I got it to work after a fair amount of back-and-forth. It was ultimately only able to organize the image files into folders by the date and time they were created (the EXIF capture date), which made further organization easier.

I then used Hermes Agent to do the same thing, and it prompted me with a decision tree of choices before embarking on its quest, which was cool to see. Despite ultimately coming to a solution, these applications are buggy and often behave unpredictably. Hermes often stops displaying my inputs in the chat but still records them (or vice versa), whereas OpenClaw requires a lot of tinkering and upkeep to keep the gateway up and running.

Both initiate what feels like heavy compute power — fans running, progress wheels spinning, and heat generated — for seemingly simple requests. Other times, responses were near-instant. In the same vein, responses from my local AI agents were often surprisingly insightful, providing inbox summaries and daily to-do lists based solely on what I gave them access to in the Chrome browser.

Also: I had high hopes for Googlebooks

For visual and creative workflows, I wanted to see how fast the Surface Laptop Ultra could generate video clips from still images using Comfy UI, and was impressed with its speed and efficiency. It can generate 10-second video clips in about 30 seconds, with 20-second video clips in about a minute, and the best part is that all of this runs locally on-device, without sending anything to the cloud. (I disconnected from Wi-Fi manually just to prove the point.)

This is where it all starts to make sense. For users who rely on AI-powered workflows, the option to run them locally on-device is extremely compelling. Privacy concerns alone are enough to justify investing in a laptop like this, but when you consider the costs of token generation and cloud compute, the Surface Laptop Ultra’s high up-front cost makes a lot more sense.

It’s here that I see the most potential for the average Windows user. This is, by far, the furthest we’ve come with this kind of technology, and it absolutely represents the future of computing. But as it stands now, many of these local models aren’t exactly user-friendly, behave erratically, and still come with inconsistent results. But we’ve never been closer to powerful, on-device models that are ready to go out of the box, and this laptop proves it.

What about for gaming?

Everyone I’ve spoken to is curious about the Surface Ultra’s gaming performance. It’s not a dedicated gaming device, but the hardware is certainly robust enough to run the latest titles, with some caveats. The first is that Nvidia’s N1X processor runs Windows on ARM architecture, so there could be compatibility issues with certain titles, as Windows runs them via emulation.

I ran “Gears of War: E-Day” on the Surface Laptop Ultra with DLSS settings enabled and again, wasn’t blown away by the performance. In fact, I had to reduce the settings substantially to get around 50fps. “Marvel Rivals”, a game that’s native for Arm, still only got around 60fps in my testing on average, and that’s not with graphics at max. Still, other less-demanding games in my Steam library looked gorgeous on the 15-inch display, like “Civilization VIII” and “No Man’s Sky.”

Also: I tried replacing Claude Code with a local AI stack that’s completely free – how it went

All of this is to say that the Surface Laptop Ultra is not necessarily a gaming laptop, but the hardware is robust enough to handle many games outside the top tier just fine, as long as you’re aware of any trade-offs that come with Arm emulation.

Final thoughts

The Microsoft Surface Laptop Ultra is a powerful, premium package, but it’s a very exclusive device. This is a laptop for power users, AI enthusiasts, and developers looking for the bleeding edge of local AI compute. To get the most out of it, you should fall into one of those camps. If you’re just using AI to summarize documents and create pitch decks, it’s overkill. Even the tasks that I experimented with admittedly only scratch the surface of what this laptop can do.

Much of this device’s potential is still locked behind usability concerns, though. I realize agentic AI isn’t exactly aimed at the mainstream consumer, but isn’t that where developers like Microsoft and Nvidia want it to be? Isn’t that why data centers and AI infrastructure are being built at breakneck speed across the country? If you want “an agent on every PC,” it’ll have to get past the feeling that we’re in a beta test.

As a single-minded device for developers and pros at the frontier of agentic AI, it’s naturally going to have trouble appealing to the mainstream consumer. However, it’s obvious this is the direction Windows PCs are going, and they’re going there fast. Even a year ago, much of what this laptop can do wasn’t possible. In another year, maybe all of these concerns will be utterly moot with what we’ll be able to do.