Nvidia's new RTX Spark laptops are sold on local AI, yet the most useful thing one tester did with a local agent was prove that the cloud is faster. That is the uncomfortable shape of this launch. The hardware is handsome, the Windows on Arm story has quietly stopped being embarrassing, and gaming works better than anyone had a right to expect. The headline feature, meanwhile, still behaves like a developer preview that happens to cost up to $7,000.
What follows combines Microsoft's account of how it prepared Windows for the chip with a week of hands-on impressions from three testers using five laptops.
What you are actually buying
RTX Spark is Nvidia's Arm-based system-on-chip for laptops and small desktops, and the laptops carry the N1X processor. It pairs Arm CPU cores with Blackwell graphics and a big pool of unified memory, the idea being that sizable AI models can live on a thin machine. The laptops look like ordinary Windows machines, but the chip inside is not the x64 design from AMD and Intel.
The tested configurations all used the full N1X, with 20 CPU cores and 6,144 CUDA cores. A cheaper variant has 18 cores and 5,120 CUDA cores.
| Laptop | Memory tested | Notes |
|---|---|---|
| Microsoft Surface Laptop Ultra | Up to 128GB | The flagship placement, mini-LED display |
| Asus ProArt P16 | 64GB and 128GB | 99.9 watt-hour battery, OLED |
| Lenovo Yoga 9n | 64GB | 2-in-1, 16-inch 120Hz OLED |
| HP OmniBook Ultra | 32GB | Nearly five pounds |
Prices run from $3,199.99 for the HP OmniBook to $6,999.99 for the Asus ProArt. The cheapest entry points, at $2,599.99, are the Surface Laptop Ultra or a smaller Asus ProArt P14 with 24GB of RAM, both using the 18-core chip.
The target buyer is a developer, coder or content creator who might also want local AI and some gaming with the latest Nvidia tricks, DLSS 5 included. Nobody involved pretended this was a mass-market purchase.
Microsoft did a lot of homework
A chip like this only works if Windows treats it as a first-class citizen, and Microsoft says it spent considerable effort on that. Peter Dawoud, a principal program manager at Microsoft, walked through the main areas.
A faster path to agents. On first boot, RTX Spark laptops show a new Get Started experience. According to Dawoud, a user can find an agent, install it and reach it in only two clicks. Agents are installed and updated through the Windows Store, so people do not have to chase down dependencies by hand. Hermes, Perplexity and OpenClaw are the first three officially supported agents, with others on the way. Microsoft says it will keep adding features so agents can do more through Windows.
Emulation tuned per chip. RTX Spark laptops rely on Microsoft's Prism emulator for apps with no native Arm build. Dawoud's point was that emulation is not generic. Each core is different, and translating x64 code to a given core has to be tuned for it, so Microsoft optimized for the RTX Spark microarchitecture. Microsoft also ships a Prism update every month to improve compatibility.
There is a history here. Emulation was a sore spot on Windows on Arm for more than a decade, and Qualcomm's first Snapdragon X laptops a couple of years ago pushed developers toward native Arm apps and pushed Microsoft toward a better emulator. RTX Spark inherits both.
Memory management. The Surface Laptop Ultra can ship with up to 128GB of unified memory, enough to load huge models. Dawoud said Microsoft invested heavily in the Unified Memory Architecture so users can load as much as possible, while holding back a little so the system and its apps stay responsive. The work covered power and performance tuning, battery life, and the ability to play a game while a video renders in the background.
There is also a slider in Nvidia's Windows app that lets users decide how much unified memory goes to the GPU side. It matters most on the 24GB and 32GB models, where every gigabyte is contested.
Does it just work?
Mostly, yes, and that is the quiet headline. All three testers said Windows on Arm is no longer the compromise it was a few years ago.
One tester noted that Adobe has had native Arm versions of Photoshop and Premiere for a while, and that Lightroom Classic, still an x86 app, ran without any visible problem. His verdict was that you would need specialized x86-only software, or a desire to play any game without risk, for Windows on Arm to be a deal-breaker.
Another made an RTX Spark his daily driver with Signal, Discord, Audacity, CapCut, Evernote, Irfanview, Slack, ScreenToGif and more. Nothing failed outright, and he never had to think about whether an app was Arm or x86. There were small irritations. Apps sometimes loaded slowly or needed a second try, Evernote notes vanished a few times after alt-tabbing until a force refresh fixed it, and apps ignored his system dark mode preference.
The third tester, who has used a Qualcomm Windows on Arm laptop daily for two years, said he can no longer tell whether an app is native. His expectation is that Nvidia's arrival will improve things further for creative apps and gaming.
Can it game?
This is where RTX Spark differs most from earlier Arm laptops. Qualcomm's Snapdragon chips never had the GPU muscle to make gaming real. Nvidia's graphics hardware brute-forces its way through the emulation layer, and nearly every game one tester tried ran emulated.
The anti-cheat problem
A large share of modern multiplayer games refuse to run without anti-cheat software, and that software had to be taught about Arm. Some of that work began with Qualcomm but sped up once Nvidia got involved. Per Microsoft's account, these games will work out of the box:
- Fortnite
- Marvel Rivals
- Arc Raiders
- Counter-Strike 2
- Dota 2
Others are on a calendar:
- Call of Duty: Activision's Ricochet anti-cheat will support RTX Spark at some point in 2027.
- Battlefield: EA has committed to getting its Javelin anti-cheat working in early 2027.
- PUBG: expected in the first half of 2027.
- League of Legends and Valorant: Riot Games has committed, with no date given.
Microsoft is also bringing its advanced shader delivery system to RTX Spark. The Xbox app can precompile shaders during a download, the same experience as on Xbox Ally devices, so there is no waiting after you click launch.
What the frame rates look like
The tester who has a desktop with an RTX 5090 said medium settings at 1440p with DLSS performance will get you a stable 60fps in most games. On a 15-inch display he did not mind the look. His numbers, at medium settings with DLSS performance:
- Black Myth: Wukong: 70fps average at native 3270 x 2180
- Cyberpunk 2077: 66fps average
- Forza Horizon 6: 57fps average
Another tester got smooth 60fps in Returnal (1440p, high settings), Expedition 33 (DLSS performance upscaled to 1200p, high), Jedi Survivor (the same upscale, medium) and Shadow of the Tomb Raider (native 1800p, highest settings).
Then the CPU showed up. In Helldivers 2, even a DLSS-upscaled 1080p did not guarantee 60fps in the hardest missions, because too many enemies crowded the screen. Star Wars Zero Company was worse. Changing graphics or resolution made no difference, and the framerate sat between 37fps and 39fps in battle. Even old games had quirks. Max Payne 2, from 2003, dipped to 50fps in places, and UFO 50 had a few hitches.
This fits Microsoft's own observation that CPUs often dictate frame pacing, while the GPU does the rendering. A strong GPU cannot fully rescue a CPU that is working through an emulation layer.
Not a gaming laptop, by its makers' own admission
Microsoft and Nvidia maintain these should not be considered gaming machines. The testers largely agree. One found the chip "no graphical magician" after needing to drop a 4K Asus panel to 2560 x 1600 and High settings in Big Walk, a co-op puzzle game, to stay near 60fps. His estimate is that the chip lands between a laptop RTX 5070 and a laptop RTX 5060, even though Nvidia claims it is like an RTX 5070 Laptop GPU in an SoC. For gamers first, he concluded, money goes much further on a traditional Nvidia discrete GPU, and cheaper desktops and laptops can beat the N1X.
One tester called it more than good enough for gaming on the go, comparing it to roughly MacBook Pro level, but not as good as a dedicated gaming laptop. Heat and noise come with it, since maxing out the fans on a thin machine means loud and hot, louder than a good desktop or handheld.
VR is shaky. A Titanfall 2 VR mod worked at first and lost stability after a few minutes. The driver for the Steam Frame wireless adapter would not install, and Steam's streaming audio solution failed too, so sound came out of the laptop rather than the headset.
Battery life: big packs, modest returns
The launch laptops carry some of the largest batteries you can take on a plane without questions. That does not turn into endurance.
On the Asus ProArt P16 (99.9 watt-hour), a tester doing ordinary work in a web browser at 50 percent brightness got a best of 5 hours 40 minutes, including an hour-long video call. A day with two hours of CapCut editing gave about 5 hours, and a third drain also took 5 hours.
The power draw tells the story:
- Web work: 13 to 17 watts on default settings
- Light video work: 20 to 30 watts
- CapCut export: up to 50 watts, briefly
- Indiana Jones and the Great Circle: up to 120 watts
Under that gaming load the battery could not keep up. At 30 percent charge, performance dropped sharply as power fell to 60 watts. On the bright side, the Asus charges quickly from a 140W USB-C PD source, reaching 50 percent in 20 minutes, 70 percent 15 minutes later, and 95 percent in 1 hour 15 minutes in total.
A second tester, running benchmarks mostly plugged in, called battery life solid anecdotally but said a full eight-hour workday is unlikely unless you are careful. Run a heavy local AI model and you will not make it to lunch. He plans standard rundown tests later.
Local AI: the headline feature, for now a beta
This is the part the platform is built around, and it is the part that disappointed most.
The setup works, with an asterisk. One tester found Hermes mostly a two-click affair, but OpenClaw kept complaining that a local gateway had not been installed properly, and it took an uninstall and reinstall to fix. Another said the install experience flashes up command prompts and Visual Studio, which felt like a beta aimed at developers. A third found that after the easy install you are "kind of on your own." On systems with 32GB of RAM or less, Hermes says there is not enough memory for its recommended local Qwen models and either routes you to the cloud or leaves you to find a smaller model.
The measured results were not flattering.
The news briefing. One tester asked Hermes, running Qwen3.8-27B, to build a daily news briefing like the one in Nvidia's on-stage demo. It took 10 minutes and 10 percent of a 99.9 watt-hour battery, while hogging the GPU and heating the laptop. Claude (Opus 5.5) did it in half the time, with more detail and better sourcing. A clean reinstall of Hermes, to rule out misconfiguration, made things worse: the prompt took longer and used 60 percent of the battery. He conceded that the local tokens were free, that he would not trust Claude with his email, and that a smaller model might have done better. But he used what the program recommended.
The image job. Another tester spent 30 percent of the Lenovo Yoga 9n's battery over 36 minutes getting a Qwen 3.8 27B model to resize five PNG product images and add white backgrounds. It cropped them wrong and cut off the products. After a couple of follow-ups it got there, and he had to grab the power cable so Hermes would not kill the laptop along the way.
The use-case problem. A lot of tasks are simply quicker on a free cloud model, without fans spinning up. The real value is leaving an agent running at full throttle overnight, the kind of work developers favor. The Microsoft-side tester said much the same, expecting to use local AI for overnight or background jobs, since large local models respond more slowly than cloud ones, with coding being the obvious use.
Privacy is the genuine upside, because emails and files stay off a cloud service. Even so, one tester regularly closed OpenClaw or Hermes rather than leave them running all day with access to local files. He is not ready to trust AI fully, local or not. He is also unconvinced by the vision, shared by Microsoft and Nvidia, of agents automating the PC. He likes editing photos and video, writing and gaming, and does not want to watch a robot do it for him.
The hardware itself is excellent
When a laptop costs this much, the makers tend to bring their best effort. The testers found great-to-excellent haptic trackpads, keyboards, screens and speakers across the line, metal builds that feel good, and OLED panels that look fantastic (the Surface uses mini-LED instead). All are reasonably thin for workstation-class machines.
- Asus ProArt: one tester's pick on hardware alone, for three USB-C ports, a very bright display, a good arrow cluster and a full SD slot. Another found the extra glossy screen and oversensitive touchpad annoying, with palms triggering accidental taps. Its speakers and webcam are mediocre, though Windows Hello and proximity sensing are excellent, and the fans produce a whoosh rather than a whine.
- HP OmniBook Ultra: the nicest-feeling chassis, but nearly five pounds and with a less useful microSD slot.
- Lenovo Yoga 9n: a sleeper favorite as the only 2-in-1, with a large stylus that works on the screen or the trackpad and a tent mode that suits gaming and video on its 16-inch 120Hz OLED. It is not as bright as the Asus.
- Surface Laptop Ultra: a MacBook Pro-like look, a magnetic USB-C port described as something Apple should have done, a haptic trackpad with subtle feedback when moving windows or aligning objects, a full SD reader and a USB-A port. The tester ran into some odd bugs but could not fault the hardware, and found the "Enhanced" color setting improved contrast over the default "Vivid." He would have preferred OLED.
Against Apple, the CPU is the weak spot
The high-memory configurations are aimed at the Pro and Max chips in Apple's MacBook Pros. On graphics, they can hang, tying the top-spec M5 Max in a Blender render. On CPU, they cannot.
One tester said nothing competes with Apple's fast cores in single-core performance, Nvidia included, and the sting is that even a 12-core M6 Mac Mini, starting at $900, competes with the 20-core N1X in multicore work. He still likes graphics at this level without a discrete GPU, since it allows thinner designs and quieter fans, as with AMD's Strix Halo, but like Strix Halo it is expensive.
A colleague thinks Qualcomm sits closer to Apple on the CPU side, helped by years of Windows on Arm experimentation and its Nuvia acquisition. By his reckoning RTX Spark took at least three years to develop, and sources told him it was meant to launch last year, so the CPU being a little disappointing is notable. Nvidia chose MediaTek for the Arm CPU, a company better known for laptop Wi-Fi hardware. His forbidden dream is a Qualcomm CPU with an Nvidia GPU, which he does not expect to see.
What is still missing from Windows
Copilot Plus PC features such as Recall are not on RTX Spark devices at launch, and it is unclear why. Adrienne Brewbaker, a senior director of Windows marketing at Microsoft, said the company is working closely with Nvidia and expects Copilot Plus PC experiences to reach Windows Insiders first, later this year. Full availability looks like next year.
Microsoft is also overhauling Windows Search for speed, with actions like toggling Bluetooth from the search bar. It effectively turns search into a launcher, which one tester called a much-needed improvement.
The verdicts
The three testers landed in similar territory.
- Interest, not conversion. One expects RTX Spark to raise interest in Windows on Arm for creative work and gaming, and sees a gap for a cheaper gaming-focused version, perhaps the rumored N1 chips.
- Nice but too pricey. Another found the laptops too hot and loud when doing the things that would justify the price, and cares more about what they stand for than owning one.
- Wait for the second generation. The third wants lower-priced options and a second generation with better CPU performance to close the gap with Apple. He is most excited to try the base Surface Laptop Ultra or Asus ProArt P14 with 24GB against the entry-level 14-inch M5 MacBook Pro, which now starts at two thousand dollars. He also hopes the platform pushes Arm-compatible games forward and maybe eases porting to macOS.
All three note the backdrop: RAM and storage prices pushed up by AI demand. One wonders what these laptops would cost without it, and admits he has no idea who will spend up to $7,000 while PC sales trend down. It is cool they exist, he adds.
Where the accounts differ
The two accounts mostly agree, including on the shared gaming results. One small mismatch is the install experience, which Microsoft describes as two clicks while one tester called it one-click for Hermes Agent, OpenClaw and Perplexity. Another found Hermes mostly two clicks and OpenClaw troublesome, and a third saw command prompts and Visual Studio flashing up. The experience seems to vary by agent and by machine.
The takeaway
RTX Spark does something Windows on Arm could not do before. It runs the apps, plays the games and looks good doing it. What it has not yet done is justify the premium with the feature it was built around. For now, the best argument for a local AI laptop is that it will happily work overnight, which is also the best argument for a desk lamp with a timer.