Related: this hands-on follows our coverage of the chip and its ecosystem: RTX Spark N1X power specs, the launch software ecosystem and which Surface Laptop Ultra configuration to buy for local AI.
Nvidia's RTX Spark N1X laptops are in reviewers' hands, and the first multi-laptop verdict is more measured than the launch hype. In "A week with RTX Spark", published October 16, 2026, three Verge writers (Antonio G. Di Benedetto, Sean Hollister and Tom Warren) tested five configurations from Microsoft, Asus, HP and Lenovo. Their summary: an "exciting new challenger, but exclusively for the high-end."
This is a hands-on of impressions, not a scored review, and the Verge says so. We summarize the findings that matter to AI builders, attribute each to the reviewer, and note where Nvidia's own claims differ. For the official software story, see Nvidia's blog on local AI for RTX Spark laptops.
TL;DR
| Question | Verge finding |
|---|---|
| What was tested? | Surface Laptop Ultra (128GB), Asus ProArt P16 (64GB and 128GB), Lenovo Yoga 9n (64GB), HP OmniBook Ultra (32GB) |
| Chip | N1X with 20 CPU cores and 6,144 CUDA cores; a lower tier has 18 cores and 5,120 CUDA cores |
| Price range | $2,599.99 to $6,999.99 |
| Windows on Arm? | Largely seamless; reviewers rarely noticed native versus emulated |
| Gaming | About 60fps at medium, 1440p, DLSS performance in most games; CPU-bound games struggle |
| Battery | Around 5 to 5.7 hours of real work on the best case; heavy tasks drain far faster |
| Local AI | Works, but slow, hot and battery-hungry compared with cloud models |
Does Windows on Arm just work now?
Mostly yes. Di Benedetto says Windows on Arm is "no longer a wholly subpar experience," crediting years of developer attention driven by Qualcomm's Snapdragon X chips and Microsoft's Prism emulation. Adobe has native Photoshop and Premiere, and Lightroom Classic ran fine under emulation. Hollister made an RTX Spark his daily driver with Signal, Discord, Audacity, CapCut, Slack and more and "never once needed to think" about architecture, though Evernote notes vanished once after alt-tabbing and apps ignored his dark-mode setting. Warren, who has used a Qualcomm Windows on Arm laptop for two years, said he can no longer tell whether an app is native.
The real hitches came from gaming and AI, and from peripherals: Hollister could not install the driver for Valve's Steam Frame wireless adapter or Steam's streaming audio on the Spark.
How does it perform in games?
Nvidia pitches RTX 5070 Laptop GPU class graphics on a chip. The reviewers' results are a bit lower. Warren found medium settings at 1440p with DLSS performance gave a stable 60fps in most games: Black Myth: Wukong averaged 70fps at native 3270 x 2180, Cyberpunk 2077 66fps and Forza Horizon 6 57fps. Di Benedetto, who measured against an RTX 5070 laptop and an M5 Max MacBook Pro, put the chip "somewhere between" an RTX 5070 and an RTX 5060 Laptop GPU. In his table for Cyberpunk 2077 at Ultra settings, the Asus ProArt P16 scored 41fps and the HP OmniBook Ultra 46fps, versus 57fps for a 2025 Framework Laptop 16 with an RTX 5070 and 56fps for the MacBook Pro.
The CPU is the limit in some titles. Hollister saw Helldivers 2 dip below 60fps on the hardest missions even at 1080p with DLSS, and Star Wars Zero Company capped between 37 and 39fps regardless of graphics settings. The verdict from the group: fine for gaming on the go, not a replacement for a gaming laptop, and loud and hot when pushed.
How long does the battery last?
Not all day. Hollister's best run on the Asus ProArt P16 with a 99.9 watt-hour pack was 5 hours 40 minutes of web work at 50 percent brightness, and other runs landed near 5 hours. Power draw was 13 to 17 watts for web work, 20 to 30 watts with light video work, around 50 watts during a CapCut export, and up to 120 watts in Indiana Jones and the Great Circle, where performance dropped at 30 percent charge as the battery could no longer feed the load. Charging is fast on a 140W USB-C source: 20 minutes to 50 percent, 1 hour 15 minutes to 95 percent. Di Benedetto adds that you will not get through a full eight-hour day, and that running a local model "won't even make it to lunch."
Is local AI useful on these laptops yet?
This is the section most readers of explainx.ai will care about, and the answer is a qualified no. Warren calls it "a beta aimed at developers right now," with agent installs flashing command prompts and Visual Studio, and fans spinning up for basic queries. He found few compelling uses, because free cloud models do many tasks faster, and sees the value in leaving an agent running overnight and in keeping emails and files off a cloud service.
Hollister ran Hermes with Qwen3.8-27B to build a daily news briefing like Nvidia's stage demo. It took 10 minutes and 10 percent of his battery, while Claude Opus 5.5 finished in about half the time with more detail and better sourcing, and the laptop stayed hot and sluggish while Hermes was open. A clean reinstall took longer and used 60 percent of the battery. Di Benedetto burned 30 percent of the Lenovo Yoga 9n's battery in 36 minutes to have Qwen3.8-27B resize five product images, which it first cropped wrongly, and he had to plug in to stop Hermes from killing the machine. On 32GB systems, Hermes simply says there is not enough memory for its recommended local Qwen models.
The reviewers also raise trust: Warren says he regularly closes OpenClaw or Hermes rather than let them run all day with access to local files. This is consistent with our own Verge local AI diary lessons and with the caution we gave about agent permissions and Windows local-cloud hybrid agents.
Is the hardware itself good?
Yes. Di Benedetto says the line is "awash in great-to-excellent haptic trackpads, keyboards, screens, and speakers," with metal builds and OLED displays. He would choose the Asus ProArt for its three USB-C ports, bright display and full SD slot, finds the HP the nicest feeling but heaviest at nearly five pounds, and calls the Lenovo 2-in-1 his sleeper favorite. Hollister dislikes the P16's glossy screen and oversensitive touchpad but praises Windows Hello and proximity sensing. Warren found some bugs on the Surface Laptop Ultra but no fault in the hardware.
What does this mean for what you buy?
Putting the reports together, a practical decision guide:
- Buy for the memory, not the hype. Local models are memory-bound. The 128GB configurations are the ones that can hold large models; 32GB systems cannot run the default Qwen models in Hermes. Our Surface Laptop Ultra configuration guide works through this.
- Treat local AI as a plugged-in workflow. Overnight agents and private data are the use cases; quick questions are faster in the cloud.
- Do not buy it as a gaming laptop. A cheaper laptop with a discrete Nvidia GPU beats it, per Di Benedetto.
- Compare against a desktop alternative. Warren and Di Benedetto both use other machines for heavier work, such as a Mac Studio, and our Surface RTX Spark dev box coverage shows the desktop form factor.
- Check your software. Specialized x86-only apps remain the main reason to avoid Windows on Arm.
How should you test one before keeping it?
If you buy an RTX Spark laptop for local AI, spend the return window on measurements the hands-on could not provide. Run the model you actually plan to use, not a demo prompt, and record three numbers: tokens per second for generation, time to first token on a long prompt, and battery percentage consumed per hour of agent activity both plugged in and unplugged. The Verge's anecdotes suggest the third number is the one that surprises people, with figures ranging from 10 percent of a battery for a ten-minute briefing to 60 percent on a repeat run.
Next, test thermals. Hollister noticed the laptop staying hot and sluggish as long as the agent app stayed open, which suggests background residency matters as much as active generation. Check whether closing the agent restores normal performance, and whether the fan profile can be tamed without throttling.
Finally, compare with the cloud on your own tasks. The reviewers' pattern, that a free or paid cloud model finishes sooner with better sourcing, may not hold for your workload, particularly when privacy rules forbid sending data out. The honest conclusion is that these machines make local inference possible and private, not faster or cheaper per task than the cloud, and the purchase only makes sense when privacy, offline use or overnight batch work is the reason.
What the Verge has not tested yet
The writers stress limits. They had a week, tested four of the six manufacturers' laptops, and ran only partial battery rundown tests (Di Benedetto expects to do more). Nothing here measures model throughput in tokens per second, which matters for builders; for that, see how llama.cpp runs on the platform in our Llama Windows coverage. Treat these impressions as early, and expect software to improve, because much of the local AI friction the reviewers describe is installer and driver polish rather than silicon.
Related reading
- RTX Spark N1X power specs
- RTX Spark launch software ecosystem
- Surface Laptop Ultra: which config to buy
- Surface RTX Spark dev box
- Llama Windows: llama.cpp tray app
- The Verge local AI diary: lessons
- Windows hybrid intelligence for local and cloud agents
Sources: The Verge: A week with RTX Spark · Nvidia blog
Accurate as of October 16, 2026. Findings are the Verge reviewers' own impressions from about a week of testing, not independent measurements by explainx.ai.