Bottom line. NVIDIA has moved its Grace-Blackwell "Spark" concept from a developer desktop into mainstream Windows laptops. The design carries one load: a large pool of unified memory, meant to let models and agents run on the device instead of in the cloud. The memory ceiling is the best-corroborated fact in the record. Almost everything else, including pricing, supply and independent performance, is unsettled or thinly measured. The memory figure is solid. The claims resting on it have not yet been tested.
What the record establishes
NVIDIA announced RTX Spark on October 16, 2026 26,45, and laptops went on sale the same day 13,15. The platform is described as built to run LLMs locally rather than in the cloud 15. It also extends NVIDIA's reach from discrete GPUs into the PC system itself 16.
The strongest single claim is the memory ceiling. The platform supports up to 128GB of unified memory, and 16 sources assert it 1,2,3,4,6,7,8,9,10,11,18,26,32,39,44. The silicon is described as a superchip integrating a Grace CPU, a Blackwell GPU and unified memory 2,10,39,44. The CPU cores are Arm-based 37, and the platform is described as using NVLink-C2C 11,38. CPU and GPU share one memory space 34. NVIDIA says this removes the memory-transfer bottlenecks that come with PCIe 37. That last point is a vendor claim. The material offers no independent measurement of it. Three sources say MediaTek collaborated on the custom CPU design 33,39,44.
Distribution is also well supported. Laptops are offered by ASUS, Dell, HP, Lenovo, Microsoft and MSI 19,44. Named models include the ASUS ProArt P16, backed by six sources 20,21,22,23,25,28, the Dell XPS 16 33,39 and the MSI Prestige N16 33,39. Microsoft's Surface Laptop Ultra also carries the chip 12. Six OEMs on day one suggests the platform has real channel support. It says nothing about whether the machines perform as claimed.
Tiering, pricing and supply
The product is segmented by core count and memory. The 18-core part is limited to 24GB and 32GB 28. The 20-core part spans 24GB to 128GB 28. Both run memory at 9,600 MT/s 28. Only the 20-core part reaches the headline capacity. The top 20-core GPU has 6,144 CUDA cores 28, the same count as an RTX 5070 42.
Supply at the top of the range looks tight. The Surface Laptop Ultra's cut-down chip carries 24GB 36, and its 128GB variant was marked out of stock 30. That rests on a single report, so it is a signal and not a finding.
Pricing is the least settled area, and the sources disagree:
- Six sources put laptops at $2,399 and up 19,24,29,32,33.
- Others say prices start at $2,599 18,39, and The Register puts retail configurations at around $2,600 42.
- The cheapest machine one report found was the ASUS ProArt P14 with 24GB at $2,599.99 29.
- Top-tier maximum-memory configurations exceed $6,000 42.
- One commenter cited a 128GB HP SKU at $3,800 41. Others attribute that figure to a specific SKU with a discount coupon 41.
NVIDIA has not confirmed official retail pricing across the lineup 15, and pricing may vary by region 39. The $2,399 versus $2,599 gap is probably a difference between entry configurations and reporting. That is interpretation, and the material does not resolve it. A buyer should treat the floor as roughly $2,400 to $2,600 and the 128GB configurations as a separate, costlier class.
Performance claims against the evidence
NVIDIA's claims are application-specific. It cites Adobe acceleration of up to 2x in Photoshop, Premiere and Substance 3D 15, live from October 16 15. Photoshop and Premiere have native beta support at launch 44. Substance 3D support is described as coming soon 33,44, so the 2x claim is not yet fully deliverable across all three applications. NVIDIA also claims 8K 4:2:2 editing and 90GB 3D scenes while unplugged 44, and AAA gaming at 1440p above 100 fps 44. It says RTX Spark is more capable than the Snapdragon X Elite and Steam Deck 36.
Independent grounding is thin and mixed. The headline "1 petaflop" figure refers to FP4 AI compute. It does not indicate game frame rates or Windows program speed 30. The few comparisons come from single commenters, and they disagree. One finds ARM-native 3DMark on par with an RTX 5070 Mobile 41. Another finds Cyberpunk about 25% behind an RTX 5070 41.
Memory bandwidth is a plausible explanation for the gap. The RTX 5070 has more than twice the bandwidth of the GB10 GPU 31, and DGX Spark's bandwidth is 273 GB/s 27. The material does not establish that bandwidth causes the Cyberpunk shortfall. Still, the pattern fits a platform tuned for capacity over throughput. Capacity decides whether a large model fits in memory. Bandwidth decides how fast it runs once loaded. The sources give strong evidence for the first and little for the second. Some users also flag driver maturity and the AI-focused niche as areas to improve 41.
None of the supplied evidence reports sustained-load behaviour, such as thermals, throttling or battery draw under the "unplugged" claims. A 128GB laptop with unified memory is a structure whose limits only show under load. Until measured results exist, NVIDIA's unplugged figures are specifications and not findings.
The software bet
NVIDIA appears to understand that capacity is worth little without a toolchain. It cites the first PyTorch-CUDA build for Windows on Arm 13,44, a native Python 3.15 build 13 and support for 97% of the top 15,000 Python packages 39,44. WSL is supported 33. Microsoft reworked Windows separately for these laptops 34. More than 130 accelerated applications are supported 15. NVIDIA names anti-cheat support for titles including Fortnite and Roblox 36.
On the AI side, Nemotron 3.5 Bolt 74B-A9B 24GB is built for these laptops 14. Four sources say NVIDIA PAIR lets an RTX Spark serve inference to Windows, Linux and macOS machines on a home network 33,44. Local processing is pitched as keeping code and context on the device 44. This is the only privacy argument in the material, and it is NVIDIA's own framing. Developers are offered a path to deploy across more than 150 million RTX AI PCs and cloud GPUs 44. That is a lock-in mechanism as much as a convenience, since CUDA-built work moves easily within NVIDIA's ecosystem.
Interpretation: capacity as the wedge
What follows is inference, not source fact. The evidence is consistent with NVIDIA using memory capacity as its wedge into the laptop. Reports say the launch challenged Apple Silicon's memory-capacity lead 17, and the platform is compared in class to AMD's Ryzen AI Max 31. The CUDA-centred software stack and the laptop-to-cloud path may matter as much as the silicon. NVIDIA says Spark is a family and lower-cost versions are in the works 36. That would matter, because the current price floor sits above $2,300.
The risks are structural. Pricing rises steeply at 128GB. Third-party performance data is anecdotal. NVIDIA itself sees memory as an accelerator bottleneck through early 2028 40. A separate report says an add-in-board partner can no longer order the 32GB GDDR7 needed for RTX 5090 cards 47. Together these suggest memory supply could constrain high-end configurations. The supplied material does not link the two points to RTX Spark directly, so this is a hypothesis.
Several details also conflict or lack support:
- An older-dated claim says the superchip would launch in the Fall of 2025 5,43,46. That conflicts with the October 2026 launch.
- A test ASUS ProArt P16 unit had 64GB 35, while the retail tiers are 24GB, 32GB and 128GB 35.
- A Malaysian listing offers a 48GB configuration 28.
The first is most likely a stale projection. The other two may reflect review hardware or regional variants. The material does not say.
What to do with this
The platform's capacity claim is well supported. Its speed claims are not yet. For a buyer or analyst, the following follows from the evidence:
- Treat memory as the product. The 128GB ceiling is corroborated by 16 sources 1,2,3,4,6,7,8,9,10,11,18,26,32,39,44. Only the 20-core part reaches it 28.
- Verify price and stock per configuration. Reported entry prices differ by $200 19,24,29,32,33,39, and a 128GB variant was reported out of stock 30.
- Ask for sustained-load data and bandwidth-bound results. The petaflop figure is FP4 AI compute only 30. Early gaming comparisons are anecdotal 41.
- Watch the software roadmap. Substance 3D support remains pending 33,44, and the Windows-on-Arm toolchain is the main evidence of ecosystem progress 13,44.
The design is credible on capacity. Whether it holds up under sustained load, at a price buyers accept, is a question for independent measurement.