I tested Microsoft’s $2,600 Surface Laptop Ultra with RTX Spark: 8 things I learned

Does “free tokens” mean anything to you? If so, read on. What about quality Windows on Arm gaming? Get in line. Solid productivity? You’re in the right place. The Surface Laptop Ultra and its RTX Spark brethren are here.

I tested Microsoft’s Surface Laptop Ultra for about a week, pushing it to its limits. You can read my full review here, but be warned: it clocks in at about 5,000 words, only because I’m an obsessive productivity nerd. If I want to know what makes it tick, why shouldn’t you?

My editors suggested something a bit more succinct, summing up the high points of the review. So here’s the story of the Surface Laptop Ultra and its brethren: a surprisingly average CPU, paired with a terrific GPU and good battery life, but with all the qualms Windows on Arm brings.

I thought, oh, Arm — haven’t all the problems been solved with Qualcomm’s Snapdragon chips? Apparently not! I was forced to work with a beta version of Adobe Premiere, so there’s development work to be done.

Here’s what you need to know about the Surface Laptop Ultra.

1.) The base price is $2,599. It tops out at $5,899.99

It’s the K-shaped economy in a nutshell. My family produced gasps when I told them the price was $2,599 price. When I explained that was the cheapest model, they gasped again. But for the puffy vest crowd swarming San Francisco for TechCrunch Disrupt, $6,000 or so is just the cost of doing business, especially if they can do other productive activities as AI agents code, chat, and communicate.

Still, one of the gaming laptops I tested against — the ROG Strix SCAR 16, with a Core Ultra 9 275HX, RTX 5090, and 2TB of storage — costs “just” $4,499.99. But Microsoft will give you $1,000 for your MacBook Pro as a trade-in!

2.) Its chip is already a year old… kind of

Project Digits, which Nvidia launched in January 2025, shipped as the DGX Spark workstation a year ago. DGX Spark, like the RTX Spark, combined a CPU (co-designed with Mediatek) with with an RTX “Blackwell” GPU (designed wholly by Nvidia). The workstation ran Linux. This runs Windows.

Tried and true? Outdated? You make the call. But it does keep up with the big boys, generally.

3.) Single-core CPU performance is half that of the Mac

Modern games take advantage of everything inside the PC. But the operating system needs to run on relatively ancient PCs, and most of Windows seems to be clustered together in just a few instruction threads. Put another way, some of your day-to-day PC experience depends on how fast each individual processor core runs.

In the first day or so with the Surface Laptop Ultra, it felt a bit laggy. I wondered why. Even now the Start menu opens and closes responsively, but switching between menu items in the Settings menu is noticeably slow. Single-core CPU performance matters and the N1X chip powering the Surface Laptop Ultra stinks. That’s indicated in the chart below, which is hierarchically sorted by multicore performance, not single-core.

Macs feel faster because of the tight integration between Apple’s operating system and chips. Apple Silicon is also far beyond the Windows world. The N1X’s score (96) is way behind Apple’s M6 score (220), according to CPU Monkey.

New MacBook Pros are rumored to be launched later this month, too. In terms of responsiveness, the Surface Laptop Ultra will never keep up.

4.) Graphics performance is very impressive

There are many ways to measure laptop graphics, but I like ComfyUI as a metric. It’s an application for generating text-to-video output — AI slop. However, the app includes a built-in benchmark that averages the time for a number of representative runs, including a warmup.

I took several gaming laptops (specs listed below) and ran that benchmark on them. Acer’s laptop is the most recent with a creator-class RTX 5060 inside it, but the Laptop Ultra is in another tier, closer to the massive Asus ROG Strix at the top of the heap.

ComfyUI’s latest version does explicitly support RTX Spark architecture, so there’s that. But optimizations like this benefit the user regardless. You can see my review for other comparisons.

5.) App incompatibility may still be a thing

When Qualcomm’s Snapdragon chip first debuted (even before the X1 Elite), some applications written for X86 chips from AMD and Intel wouldn’t run on Arm. Some of that was alleviated by Microsoft’s Prism emulator in Windows, which translated X86 instructions to Arm native code. But just like a conversation with someone who doesn’t speak the same language as you, translation slowed the process down.

I was surprised to find that the RTX Spark family of laptops may have the same experience, though this time from a GPU perspective. Adobe Premiere needed a special beta to run. Without it, Premiere complained of an unsupported GPU. ComfyUI specifically supported the RTX Spark. I don’t think that games will have any issues with the RTX Spark, but I’m not entirely certain we’ve put this issue to bed, either.

6.) It’s a perfectly functional productivity laptop

The Surface Laptop Ultra isn’t just for AI nerds. The laptop functioned perfectly well as an ordinary, day-to-day 15-inch laptop. It is more than four pounds, however. While it’s compact, it’s neither thin nor especially light. I think of it as a relative to the Surface Laptop Studio, but without the cool pull-forward screen.

Battery life was about 10 hours while performing a set benchmark simulating office work, so that’s fine.

7.) Memory matters with AI more than you’d think

Why wait all this time to discuss AI? Because I’m still convinced that the lower-end models of the Surface Laptop Ultra and its competition should be seen more as “ordinary” laptops, while the pricey, high-end models lend themselves more to AI.

The RTX Spark architecture uses shared memory, so the laptop I was given with 32GB of RAM splits that between the CPU and GPU. A gaming laptop with a dedicated GPU contains its own VRAM. That’s a subtle but significant difference!

Microsoft believes that you’ll be able to run AI (agents or LLMs) at the same time as you work elsewhere on the machine such a web browser. I did load the Hermes agent and let it do its thing while I worked elsewhere, but I don’t really know if AI performance suffered as a result. I didn’t have enough time to test this as thoroughly as I would like.

The fact is, though, that the largest AI models can demand more storage and RAM than compressed, quantized AI models. And if there’s a super-cool AI model that everyone’s talking about, you might not be able to run it if you opt for a low-end machine.

8.) The cultural aspect troubles me

The compelling reason to use local AI is you can generate “free tokens” without a cloud subscription. You don’t need to be connected to the internet. But there is a price to be paid, both in power consumption and fan noise.

I still wonder what that looks like in the real world. I caught the Surface Laptop Ultra pulling about 130 watts while gaming, and I suspect that intensive LLM work will require the same.

Imagine traveling to Asia, seated next to someone using a Laptop Ultra. Limited connectivity, so they’ve gone to local AI. Is there enough power from the seatback to even allow this? Will the laptop fans howl while you’re trying to sleep? What will your neighborhood coffee shop be like with tech bros all maxing out their laptop cooling? It’s like a little portable data center, but right next your ear.

The big brains who think through these things never seem to understand that consumers like technology when it works for them. I just hope that ends up being the case.