Surface Laptop Ultra review: a different kind of MacBook Pro

After setting up Microsoft’s Surface Laptop Ultra, I spent five minutes just staring at it. I wasn’t surprised by its MacBook Pro-like design or blown away by its bright display. I was trying to work out what I’d actually do with a $5,900 laptop.

This isn’t just any regular Windows laptop. Microsoft has teamed up with Nvidia to build a high-end laptop to showcase Nvidia’s new Arm-based RTX Spark chips. It’s a big moment for Windows on Arm, as these chips can run demanding AI models locally, alongside the latest AAA games and creative apps. The promise is compelling, but the reality of local AI is much messier.

Getting the most out of the Surface Laptop Ultra, especially my top-of-the-line model, requires coding experience or the patience to experiment with models and AI agents until you find something that works. If that sounds like you, Microsoft might have built the Windows laptop you’ve been waiting for. If not, this might not be the MacBook Pro competitor Microsoft was hoping for.

It feels like Microsoft has finally built the hardware to rival Apple’s pro laptops, thanks to an all-aluminum design, large haptic trackpad, and high-end finish. Even the “nightfall” color is just like Apple’s space black.

There are plenty of meaningful differences, though. The 15-inch MicroLED display can run at up to 2,000 nits of peak HDR brightness, at 262 pixels per inch (3270 x 2180); the MacBook Pro peaks at 1,600 nits. That makes it the brightest display Microsoft has ever shipped on a Surface Laptop. Unlike the MacBook Pro, it’s also a 3:2 touchscreen, perfect for writing code, reading text, or editing photos.

The port selection is similar to the MacBook Pro, with one very big change. You get three USB-C ports, a USB-A port, a single HDMI 2.1b port, SD card reader, and headphone jack. Where it really differs is Microsoft’s new Magnetic Connect. It works just like the old Surface Connect, but through the USB-C port on the right-hand side. Just bring the connector close to the port, and it snaps into place. When you’re not charging, that port works just like the rest of the USB-C ports on the laptop.

Magnetic Connect is one of my favorite additions on the Surface Laptop Ultra. I love how you can reverse the magnetic USB-C charging end, attach it to the power brick, and use the regular USB-C connector at the opposite end of the cable to charge the rest of your devices.

The haptic feedback on the large trackpad is also a welcome addition. It works with things like clicks, drags, and window alignment throughout Windows 11, and the tactile feedback triggers in apps like PowerPoint as you drag and align objects or images in a deck, too. It’s subtle enough to make it feel like you’re really moving windows or interfaces around, but I do wish more apps supported it.

Microsoft has included a 1080p front-facing camera above the screen. The quality is fine for video calls, and it supports Windows Hello face authentication and Windows Studio Effects like automatic framing, eye contact, and background blur.

Not all of the Surface Laptop Ultra hardware is impressive, though. I’ve had issues where I’ve unplugged the charger at 100 percent only for Windows to then immediately report 97 percent battery life. Plugging in would put it back to 100 percent and then it would immediately drop back to 97 percent again each time I unplugged.

It seemed weird so I asked Microsoft about the behavior, and it’s designed this way to make a battery last longer. “When it’s plugged in and the battery is full, the device pauses charging rather than constantly topping off, and resumes only once the level drops to 95 percent,” says Microsoft spokesperson Moses Nsereko. “This reduces wear over time. While plugged in, the indicator shows 100 percent to indicate a full charge. When you unplug, it shows the battery’s exact level right away, which may be a few percent lower.”

I’ve run into some issues with the display, too, where either things felt slightly laggy or there was noticeable ghosting. Moving a window around rapidly sometimes left a little trail. Changing power settings from the default “best power efficiency” mode to “best performance” seemed to fix most of the laggy feeling, and the backlight ghosting seems to be less noticeable the longer you run the laptop.

I also heard some loud pops and bangs from the built-in speakers on the Surface Laptop Ultra. They were loud enough to sound like firecrackers, and I thought the battery was about to die before I realized it was just speaker interference. Rebooting the Surface Laptop Ultra solved this one, but it’s happened a couple of times now.

Speaking of rebooting, I’ve noticed that rebooting the Surface Laptop Ultra is surprisingly slow. This seems like something Microsoft could easily fix with an update, though.

Local AI models and agents

The RTX Spark that powers the Surface Laptop Ultra is designed with local AI models and agents in mind. The variant I’ve been testing has a 20-core CPU that was codesigned by Nvidia and MediaTek paired with a 6,144-CUDA-core GPU — roughly comparable to a mobile RTX 5060 — plus 128GB of unified memory and a 1TB Gen 5 SSD. It’s the top-of-the-line model at $5,899, and I think if you’re considering this particular laptop spec then you have a very specific use case in mind. The base version is $2,599, and it comes with an 18-core CPU, 5,120-CUDA -core GPU, and 24GB of unified memory.

Out of the box, Microsoft prompts Laptop Ultra owners to pick an AI agent. You can choose between Hermes, OpenClaw, and Perplexity. I tried out Hermes and OpenClaw over the past week, but I ran into issues getting OpenClaw to work initially. A reinstall fixed a gateway issue, and then I could name my agent and start using it for local AI tasks.

I configured both Hermes and OpenClaw to use the Qwen3.8-27B model, one of the more powerful local AI models. Because it uses the full power of the Surface Laptop Ultra, it will spin up its fans at full throttle and 100 percent GPU for even the most basic of queries. I asked Hermes to resize and convert some photos from PNG to JPG, a task that would take seconds using a PowerToy utility. It took minutes to complete, all while the fans were buzzing away and the laptop was getting hot.

It’s probably my fault for using such a large model for a basic task, but without testing every single local AI model it’s hard to know which one to pick for each task you need. I asked OpenClaw to build me a smart home dashboard, but it spent 30 minutes spinning up the laptop’s fans and calling various tools to eventually produce nothing but an error.

I don’t have a base Surface Laptop Ultra to test, but it’s easy to see that 24GB of unified memory will mean fewer local AI models to choose from.

I struggled to come up with many compelling use cases for the power of these chips for local AI work. A lot of tasks could simply be done far more quickly and without the laptop fans spinning up by using a free cloud model. The power here is really in leaving the agent running at full throttle to build something overnight, or run a task across large datasets that would normally cost a lot of money in cloud tokens.

I did like that I could do things locally, though, without worrying about the privacy aspects of cloud models. I fed Hermes some audio and video files and it went ahead and created a local llama.cpp server to enable audio transcription. It took some fiddling to get it working, but it could certainly save me having to install transcription apps or pay for a cloud service like Otter.

The thing I didn’t like is that these agents are far from user-friendly, and I’m surprised that Microsoft or Nvidia hasn’t developed its own local AI interface that simplifies a lot of this work. I need to spend more time with both Hermes and OpenClaw, but as someone who doesn’t code it’s clear this is all really targeted toward developers and power users.

Gaming and creative tasks

The Surface Laptop Ultra might not be an official gaming laptop, but it’s certainly capable of running far more of the latest games than Apple’s MacBook Pro. Normally I’d be nervous about Windows on Arm and gaming, given the weak GPU on the Qualcomm Snapdragon X series chips. But this is the first time we’ve seen an Arm CPU paired with a stronger Nvidia GPU, and most games run well for a laptop that’s not the size of a gaming one.

Instead of running a bunch of benchmarks at different settings, I decided to try and find the optimal settings that would allow me to play at around 60fps or more in each game. Because this is an RTX GPU, you also get full support for DLSS, Frame Generation, and ray tracing. So you can enable Frame Generation if you want to get to 120fps to match the 120Hz display.

I was able to run Black Myth: Wukong at 70fps on average at native resolution (3270 x 2180) with medium settings and DLSS performance mode enabled. I ran the same medium settings and DLSS performance in Cyberpunk 2077 and averaged 66fps, and Forza Horizon 6 managed an average of 57fps.

Most games will need to run at medium settings and DLSS performance or balanced to get around 60fps, but there are some exceptions where you’ll need to lower settings or resolution. Gears of War: E-Day averaged 40fps on medium settings and DLSS performance, and I only managed to bump this to 48fps on low settings. I also had to move Doom: The Dark Ages to low settings and DLSS performance to hit an average of 60fps.

I’m spoiled by using an RTX 5090 on a 32-inch 4K screen every day with maxed-out game settings, but I didn’t mind the look of medium settings on a 15-inch display. It felt more like dropping down to console settings, and in some games it’s hard to distinguish the texture differences on a smaller display. I’m one of those gamers who prefers to run at 60fps with lower settings rather than max out the settings and have to play at 30fps, though.

While game compatibility is good for most titles, there are some where you’re going to run into emulation issues. Horizon Zero Dawn Remastered was limited to 30fps at 1440p regardless of whether I used low or medium settings. I believe this is because the game is CPU-heavy, so a native Arm port would probably help for this type of game.

There aren’t many native Arm games, though. I tried out the new Arm version of Marvel Rivals, and it averaged 64fps with medium settings and DLSS performance and 75fps with low settings. I think we’ll only see Arm ports where there’s a lot of CPU activity involved that the Prism emulator — which translates CPU instructions from x86 to Arm — can’t handle well enough, because otherwise it’s very capable.

We may see more Arm ports for games that have anti-cheat engines. Anti-cheat engines are a lot more complicated to emulate, so developers have to create Arm-compatible versions. While games like Arc Raiders, Counter-Strike 2, Dota 2, and Fortnite all work out of the box, Nvidia and Microsoft are working with EA, Activision, and Riot Games to get titles like Battlefield, Call of Duty, Valorant, and League of Legends fully supported.

The Surface Laptop Ultra is also a very capable machine for creative and everyday tasks. I managed to export our 4K video test in two minutes and 18 seconds on the Surface Laptop Ultra using Premiere Pro. That’s around a minute behind the top-of-the-line MacBook Pro, but beats the base M5 chip and is around 30 seconds behind the M5 Pro. I did have to create this using Premiere Pro’s own export tool, rather than sending it to Media Encoder like how we usually run this test. There seems to be a bug in the beta Arm version of Media Encoder that prevented it from running the export correctly.

Where the Surface Laptop Ultra really falls behind Apple’s M5 chips is in CPU performance. Single-core and multi-threaded CPU tests like Geekbench or Cinebench show how much weaker the MediaTek CPU is against a $6,999-plus MacBook Pro 16 with an M5 Max or even an M6 Mac Mini. Even some of the GPU tests for Geekbench show the RTX Spark falling behind Apple’s MacBook Pro, and the Puget Bench Photoshop, Premiere Pro, and DaVinci Resolve are all noticeably behind the M5 Max.

I’ve also been testing the battery life of the Surface Laptop Ultra, which seems to be around seven hours if you’re using local AI models and a mix of apps and gaming. If you’re doing less intensive work, it’s averaging out at 10 hours or more. I’ll know more the longer I use the Laptop Ultra, so check back for updated battery impressions.

The Surface Laptop Ultra doesn’t really offer a clear performance advantage over the MacBook Pro outside of some AI workloads like image generation and the ability to run CUDA compute natively. It’s not transformative for the Windows world like the M1 chips were for macOS, largely because Qualcomm already supercharged Windows on Arm with its Snapdragon X series two years ago. That makes it difficult to recommend unless you’re specifically interested in running local AI models, coding, or experimenting with AI agents.

The Surface Laptop Ultra, and Nvidia’s RTX Spark in general, targets developers and power users who want a laptop that can handle those workloads alongside creative apps and the latest games. For that audience, there’s a lot to like. For everyone else, more affordable laptops can handle everyday productivity, creative work, and some light gaming without the hefty price premium.

Microsoft has come closer than ever to building a Windows alternative to the MacBook Pro. The hardware looks and feels great, and Nvidia’s GPU brings CUDA support and decent gaming capabilities to a high-end laptop design. But the Surface Laptop Ultra’s biggest selling point is also its biggest question mark: local AI. It all feels like a work in progress. If you’re ready to experiment with it, this could be a compelling glimpse of where Windows laptops could go next. If you’re not, it’s a lot of money to spend on an experiment.