Tested RTX Spark: ProArt laptop finished 120B workload in 33 minutes, versus 53 minutes for RTX 5090
The first machines with NVIDIA's RTX Spark are on sale and fully tested: the Surface Laptop Ultra was unveiled in San Francisco, goes on sale October 16 from $2,599, and is followed by an OEM wave from ASUS, Dell, HP, Lenovo and MSI. It is the first time the Windows-on-ARM market gets a genuine chip alternative — and the independent testing shows a split picture: unified memory that beats a desktop RTX 5090 card on one large-model workload, broad but not flawless game compatibility, and prices from roughly $2,600 to $7,000 that make the machines poor value for gamers.
What was actually launched
Microsoft showed off the Surface Laptop Ultra with NVIDIA's RTX Spark chip at an event in San Francisco in October 2026 (Barchart via Yahoo Finance). According to that coverage, the machine has a Blackwell GPU and up to 128 GB of unified memory, will be available from October 16 with prices starting at $2,599 — and if you pay for the developer variant, there is a compact workstation called the Surface RTX Spark Dev Box starting at $5,999. (The CPU spec is reported inconsistently across sources; see below.)
Microsoft is not alone. ASUS, Dell, HP, Lenovo and MSI are all launching Windows machines powered by RTX Spark, according to the same coverage. Microsoft itself frames the launch as part of a strategy for what the company calls "hybrid intelligence" — the idea that AI applications should be able to use compute locally or in the cloud depending on the task. That is a company statement, not an independent finding, and the underlying Microsoft announcement is not available in the source material — all launch details come via press and review coverage.
The shift: From one chip supplier to two
The structural novelty is not the chip itself, but what it represents for Windows-on-ARM. ARM chips for Windows have until now been effectively limited to Qualcomm's Snapdragon series, writes BigGo Finance in its analysis of the launch — making the RTX Spark machines a turning point that breaks the monopoly. It is worth stressing that this is BigGo's editorial interpretation, not a verified description. The article suggests — with explicit hedging — that some kind of business relationship between Microsoft and Qualcomm may have underpinned the exclusivity, but this is unverified, and the precise mechanism that ended the exclusivity is not documented in the sources.
BigGo further suggests that Windows 11 hardware will be reorganised into three tiers: non-AI PCs without AI hardware, Copilot+ PCs that handle small and medium models via NPUs, and the RTX Spark class that can run large language models locally. This too is analysis based on the source material, not an official Microsoft framing.
The architecture: SoC with unified memory and a Blackwell GPU
ZDNET's David Gewirtz got to test a $4,000 Lenovo Yoga 9n, and confirms the core of the platform (ZDNET, Oct. 26): RTX Spark is a SoC architecture with unified memory, distinguished by a powerful Blackwell RTX GPU with 6,144 cores and CUDA. The CUDA compatibility is the point for developers — the existing AI tool ecosystem is largely built around NVIDIA's software platform, and RTX Spark brings it into a portable ARM form factor. The tested Lenovo unit had 64 GB of RAM.
But there is a significant specification conflict in the source material. BigGo describes the SoC as a Blackwell GPU with a CPU of up to 10 cores based on Cortex-X925/A725 cores. ZDNET refers to a "20-core NX1 processor," and PCWorld's benchmark report refers to a 20-core ARM-based Grace CPU. The core count and CPU architecture of the chip that actually ships are thus not settled across the sources, and should not be presented as resolved here.
The benchmark: ProArt P16 versus a desktop RTX 5090
The most talked-about test is PCWorld's comparison of a $6,999 ASUS ProArt P16 against a desktop RTX 5090 card on the same real-world gpt-oss-120b inference workload (PCWorld via MSN). The result: the ProArt processed 392,181 prompt tokens and generated 79,903 completion tokens — 472,084 tokens in total — in 33 minutes and 13 seconds. The RTX 5090 card ran the same frozen workload in 52 minutes and 51 seconds, according to PCWorld's benchmark results.
It is worth taking the tester's own caveats seriously, as PCWorld Executive Editor Brad Chacos himself does: this is not a laboratory-perfect tokens-per-second figure, and the comparison is not apples-to-apples. The RTX 5090 actually generated more completion tokens in the same run than the ProArt's 79,903, so the 20-minute time win is not a raw picture of chip superiority. The mechanism driving the result is not mysterious either: the ProArt's 128 GB of unified memory lets the entire large model sit directly available to the GPU, while a desktop card has to fetch data over PCIe and hits a bandwidth cliff when the model exceeds VRAM. For this particular workload — large model, local memory — the architecture wins. It is a real and concrete advantage, but a narrow one.
Gaming: The compatibility holds, the value does not
On the gaming side, PCWorld's Michael Crider tested HP's $3,199 Omnibook 16 Ultra (PCWorld). ARM game compatibility on RTX Spark proved far better than expected — nearly every title tested ran. But the conclusion stands: launch prices of between $2,600 and $7,000 simply make the machines a bad offer for anyone primarily interested in gaming.
Performance figures like "5070-level" from the pre-launch period are NVIDIA's/the companies' own claims, not independent measurements, and should be treated as such. The independent finding is therefore: broad compatibility, but poor value per dollar compared to dedicated gaming machines.
So which machine should you choose?
Windows Central (via Cale Hunt, reproduced in Yahoo Tech) has compared the two most talked-about launch machines: the Surface Laptop Ultra, which according to Windows Central is the first RTX Spark PC it has reviewed, and the ASUS ProArt P16, which Windows Central actually rated higher than the Surface on display, cooling and design. The Surface starts cheaper, but lacks pen support. The ProArt option comes with 128 GB of RAM in the tested configuration; the same is not necessarily true of all machines.
That leads to a practical purchasing assessment: the claims about running 120-billion-parameter models locally depend entirely on the configuration. ZDNET's tested Lenovo unit topped out at 64 GB of RAM, while other machines offer 128 GB. Actual model support therefore varies with SKU and price, and "local 120B" is not a property of the platform as such, but of the most expensive configurations.
Open questions
Several things are not settled in the source material. No primary source is available: Microsoft's official announcement, NVIDIA's launch materials and any reaction from Qualcomm are missing, and all launch details are secondary via press and testing. The CPU specifications (10 vs 20 cores) are contradictory across sources, the exact mechanism behind the end of Qualcomm's exclusivity is unverified, and it is unknown whether the high-RAM configurations used in marketing reflect actual shipping units in volume.
What remains is nonetheless concrete: NVIDIA is now in the Windows-on-ARM market, with hardware that independent testing confirms delivers on large local language models via unified memory — and which is at the same time worth knowing the limits of before you buy.