On October 16, 2026, the first laptops with NVIDIA's RTX Spark chip went on sale. The same day, The Verge gave the platform its most demanding test yet: three writers, five machines, one full week of real work. Ground rules first — this is not a scored review. The Verge states plainly that no scores were awarded; what it published are impressions, experiences rather than rankings. But the notes contain one number more damning than any benchmark.
Sean Hollister asked Hermes, a local agent running Qwen3.8-27B on a 64GB Asus ProArt P16, to reproduce NVIDIA's keynote demo: a daily news briefing. It took 10 minutes and burned 10 percent of the 99.9Wh battery. Suspecting his own configuration, he reinstalled Hermes and ran the same prompt again — slower still, and this time it drained 60 percent. That is a single experience, not a benchmark, but as a cost estimate for one briefing it is more concrete than anything on the spec sheet.
The cloud comparison for the same job: Claude Opus 5.5 finished in half the time, with more detail, more sources and a better result.
The three writers' consensus compresses into one line: local AI runs, but it is slow, hot and hungry — and privacy is the only hard advantage it holds. Their verdict: "an exciting new challenger, but exclusively for the high-end." This piece unpacks what that high-end qualifier buys you, and what it does not.
1. Five Machines, One Price Band
The five test units — four models, with the ProArt P16 sent in two configurations — all run NVIDIA's RTX Spark N1X, an Arm-based chip:
| Machine | Reviewer | Memory tested | Known price |
|---|---|---|---|
| Surface Laptop Ultra (Microsoft) | Tom Warren | 128GB | from $2,599.99 |
| Asus ProArt P16, 64GB | Sean Hollister | 64GB | $6,999.99 |
| Asus ProArt P16, 128GB | Antonio G. Di Benedetto | 128GB | $6,999.99 |
| Lenovo Yoga 9n 2-in-1 | Antonio | 64GB | not listed |
| HP OmniBook Ultra | Antonio | 32GB | $3,199.99 |
The chip comes in two tiers: 20 CPU cores with 6,144 CUDA cores (Blackwell) at the top, or 18 cores with 5,120 CUDA in the entry tier found in the $2,599.99 Surface Laptop Ultra and the smaller ProArt P14 (24GB). Unified memory tops out at 128GB, and NVIDIA positions the platform against Apple's M-series and Qualcomm's Snapdragon X Elite.
The unified-memory pitch deserves unpacking, because everything else follows from it. CPU and GPU draw from one pool, so a 128GB machine can in principle load enormous models whole — that is the foundation under the "120B model with 1M-token context" claim. Traditional laptop discrete GPUs top out around 24GB of VRAM, nowhere near that class. Apple's M-series shares the same architecture logic, but its local AI ecosystem lives on the other side of an OS divide. RTX Spark is bidding to be the first Windows answer with big memory and a CUDA ecosystem attached.
Note the memory spread across the test fleet: 32GB to 128GB, a fourfold gap that later cashes out directly as "runs" versus "will not run."
2. Local AI's Three Sins: Slow, Hot, Hungry
Slow. The 10-minute briefing above is exhibit one. The same job on cloud Claude Opus 5.5 finished in half the time with more detail and better sourcing. Tom Warren, on the Surface Laptop Ultra, kept reaching the same conclusion from the other side: many tasks are faster on free cloud models, and the laptop stays cool. This is not a model-quality problem — Qwen3.8-27B is no slouch — it is a form-factor problem. A job that shifts from "seconds" to "ten minutes" stops being a tool and becomes a wait.
Hot. Local inference monopolizes the GPU and keeps the chassis warm; while Hermes ran, the machine stayed hot and sluggish. On Tom's unit, the most basic query was enough to send the fans to full speed. "Monopolizes" deserves emphasis: while a local agent works, the laptop is largely unavailable for anything else demanding. It is not a background service; it is an all-hands mobilization — the exact opposite of the "local AI, always at hand" picture on the marketing page.
Hungry. Antonio asked Qwen3.8-27B on the Lenovo Yoga 9n to put a white background behind five PNG product shots and scale them: 36 minutes, 30 percent of the battery — and the model cropped the images wrong, cutting into the merchandise itself, needing several follow-ups to get it right, all while plugged in so Hermes would not drain the battery dry. Again: a single experience, not a benchmark. But "five images, 36 minutes, 30 percent" is a conversion rate worth holding against the spec sheet.
Hollister left the quotable line, worth keeping in the original: "I guess I got the local AI tokens for 'free'." Free, provided you treat battery and time as air.
A fourth sin hides in the onboarding. Even with 128GB of memory, Antonio's verdict on the top-spec ProArt was "unfriendly to newcomers": Windows 11's Get Started offers one-click installs of Hermes Agent, OpenClaw and Perplexity — complete with ESRB ratings, oddly — and abandons you afterward. No guidance, no expectation-setting; when the fans spin up and the battery starts sliding, an ordinary user has no way to tell which program is doing the work. On 32GB or less, Hermes simply says memory is insufficient for its recommended Qwen model and sends you to the cloud or to a smaller download — which splits the unified-memory marketing from entry-tier reality right down the middle. Tom's summary: local AI "remains a developer-oriented beta," with agent installs flashing command-line and Visual Studio windows, an interface dialect that filters out most people by itself.
3. The Marketing Page Versus the Ledger
The soul of any comparison lives here: NVIDIA's claims, laid side by side with a week of receipts.
The official page: 1 petaflop of FP4 compute, up to 128GB of unified memory, 120B-parameter models with 1M-token context, all-day battery.
Against the ledger, item by item:
All-day battery. The best the three writers saw under light daily loads was 5 to 5.7 hours — clearly short of "all day" — while AI and gaming workloads burn the battery by the percentages above: 10, 30, 60.
120B plus 1M context. 128GB of unified memory does hold it, but holding and running comfortably are different things — the hands-on verdict on the 128GB machine was "unfriendly to newcomers," and machines with that much memory run $6,999.99.
1 petaflop FP4. Nobody disputes the number; it just never tells you when the fans spin up or when the battery indicator starts to slide. Paper compute and felt performance are separated by exactly the week of testing The Verge just did.
4. The Gaming Axis: Aiming at the 5070, Landing Above the 5060
NVIDIA positions the N1X against the RTX 5070 Laptop GPU; measurements land below it.
Tom Warren's Surface Laptop Ultra held 60fps in most games at 1440p medium quality with DLSS Performance: Cyberpunk 2077 at 66fps and Forza Horizon 6 at 57fps under that regime, with Black Myth: Wukong tested separately at native 3270×2180 and averaging 70fps. All figures are the Verge writers' first-hand measurements.
Antonio ran Cyberpunk 2077 across the Ultra-tier machines on one settings sheet (the piece does not tag resolution and quality per machine, so treat the sheet as internally consistent on the Verge's own terms):
| Machine | Cyberpunk 2077 |
|---|---|
| Asus ProArt P16 | 41fps |
| HP OmniBook Ultra | 46fps |
| Framework Laptop 16 (RTX 5070, 2025) | 57fps |
| Apple M5 Max MacBook Pro | 56fps |
His placement: the N1X lands between the RTX 5070 and the RTX 5060 Laptop.
A fair reading of these numbers needs the right baseline. Against a desktop gaming laptop the comparison is hopeless; against thin-and-light peers it is the opposite — touching a steady 60fps at 1440p medium is overachievement for an Arm machine without a discrete GPU. The failure is one of framing: NVIDIA drew the line at the RTX 5070, and the measurements landed above the 5060 instead. Gaming performance passes; the "5070-class" claim does not.
Then there are the CPU-bound cases, where no settings change helps: Helldivers 2 dips under 60fps on high-difficulty missions even at 1080p with DLSS, and Star Wars Zero Company holds a flat 37 to 39fps regardless of quality settings — the bottleneck is the CPU, not the GPU.
Three conclusions from the gaming axis: enough for gaming away from a desk; no replacement for a gaming laptop; loud and hot under sustained load. Thermals depend on the chassis — thick machines stay acceptable, thin ones throttle after 20 minutes of gaming, and Dynamic Boost helps without being a cure.
5. Windows on Arm: Not the Weak Point This Time
The surprise is the system layer. Windows on Arm proved largely seamless for a week: the Prism emulation layer is mature, Photoshop and Premiere run natively, Lightroom Classic runs fine under emulation, and Hollister's full daily stack — Signal, Discord, Audacity, CapCut, Slack — hit no architecture problems at all.
The dents are small: Evernote lost notes once after an alt-tab, and dark mode was ignored by one app. Peripherals are the real gap: the Steam Frame wireless adapter driver would not install, and neither would the Steam streaming audio component.
The verdict can be short: WoA is not the weak point this generation; the weak points are battery and thermals.
6. Privacy Is the Only Hard Advantage
The most honest usage pattern came from Tom Warren. He does not deny the local agent's value — overnight full-speed build tasks, and mail and files that never leave the machine — but he also does not trust a local agent with local file permissions, and closes OpenClaw and Hermes rather than let them idle in the background. The tension is real: local AI's pitch is that data never leaves, yet its permissions on your own machine are broad enough that you hesitate to let it run.
So the conclusion is not buy or skip; it is scope. If your data is sensitive, you can tolerate slow and hot, and you want the machine working overnight, privacy is the one advantage no cloud subscription can sell you. If you want speed, endurance and polish, cloud models and a gaming laptop remain better answers. The price band runs from $2,599.99 to $6,999.99, and the trio's verdict stands as delivered: exclusively for the high-end. The other side of the ledger deserves writing down too: the silicon is not the problem — the three walls of the laptop form factor, thermals, battery and onboarding, stand between paper compute and felt experience. Buyers of this first generation are not buying the present; they are buying the option on a second generation that knocks those walls down.
A note on scope within this site: our October 17 comparison put the model routes on trial, with the RTX Spark box merely serving as the test bed. This piece reverses the arrangement — the machine and the chip go on the table, and the models are just the load.
And one step further out: the open-weights Beam 501B released the same week needs four B200s to serve. Local and datacenter are the two extremes, and this generation of RTX Spark machines sits in between without a power-sipping middle rung. When the next N1X erases one of the three words — slow, hot, hungry — the arithmetic deserves redoing.
FAQ
Q1: Can an RTX Spark laptop replace a gaming laptop? A1: As a "can also game" machine, yes; as a replacement, no. NVIDIA aims it at the RTX 5070 Laptop GPU, but measurements land between the 5070 and the 5060: most games hold 60fps at 1440p medium with DLSS Performance, and Black Myth: Wukong averages 70fps at native 3270×2180 — yet Helldivers 2 dips below 60fps on high-difficulty missions at 1080p with DLSS, and Star Wars Zero Company sits at 37 to 39fps regardless of settings, a CPU bottleneck. Thin machines throttle after 20 minutes of gaming. All figures are the Verge writers' first-hand measurements.
Q2: Isn't local AI free? Why call it expensive? A2: The subscription is free; the battery is not. One briefing took 10 minutes and 10 percent of the battery, and 60 percent after a reinstall; five image edits took 36 minutes and 30 percent (single experiences, not benchmarks). Cloud Claude Opus 5.5 was twice as fast with better results. Hollister's line was ironic: local tokens are "free" — paid in battery.
Q3: Is the 128GB configuration worth the premium? A3: Depends on the job. 128GB genuinely holds a 120B model with 1M-token context, but the field verdict is "unfriendly to newcomers": one-click installs with no guidance afterward, and the fans, heat and drain are yours to manage. On 32GB or less, Hermes refuses its recommended model outright. Pay the premium only if data staying on-device is a hard requirement.
Q4: What is the battery actually like? A4: The claim is all-day battery; the best measurement under light daily loads was 5 to 5.7 hours. Under AI or gaming loads it burns by percentage: 10 for one briefing, 30 in 36 minutes for five images, 60 for the rerun. Budget for the worst case if you plan to run local AI away from an outlet.
Q5: How is Windows on Arm compatibility? A5: Largely seamless. Prism emulation is mature; Photoshop and Premiere run natively, Lightroom Classic under emulation, and a week of daily work hit no architecture problems. The gaps are scattered app quirks (Evernote losing notes after alt-tab, one app ignoring dark mode) and peripheral drivers: the Steam Frame wireless adapter and Steam streaming audio would not install.
Related reading
- Says #1, Its Own Table Says #2: Picking a 24GB Model: Says #1, Its Own Table Says #2: Picking a 24GB Model
- Half the VRAM, Nothing Runs: Five Ways Out: Half the VRAM, Nothing Runs: Five Ways Out
- Local Coding Models: From 8GB GPU to 8-GPU Node: Local Coding Models: From 8GB GPU to 8-GPU Node