NVIDIA RTX Spark laptops go on sale from $2,599, and the entry model has 24GB, not 128GB

The first Windows laptops built on NVIDIA RTX Spark went on sale on October 16. In its launch release, NVIDIA says machines from ASUS, Dell, HP, Lenovo, Microsoft and MSI are available globally through those companies' websites and major retailers, starting at $2,599, with pricing and configurations varying by region.

RTX Spark is NVIDIA's Arm-based PC chip. Three days ago we covered the software NVIDIA lined up for it; today's news is that the hardware can actually be bought, at a known price, in known models.

What is on sale

NVIDIA lists nine launch laptops:

• ASUS ProArt P14 and ProArt P16
• Dell XPS 16
• HP OmniBook Ultra and OmniBook X 14
• Lenovo Yoga Pro 9n
• Microsoft Surface Laptop Ultra
• MSI Prestige N16 Flip AI+ and Prestige N16

All of them use a single superchip that pairs a Blackwell RTX GPU with an NVIDIA Grace CPU; NVIDIA says MediaTek collaborated on the custom CPU design. The headline figures are up to 128GB of unified memory and up to 1 petaflop of FP4 AI performance. The release also mentions additional systems to follow, without naming them.

Read the "up to" carefully

Microsoft's Surface Laptop Ultra page is the clearest public example of how far the entry model sits from the headline. The laptop is listed at $2,599.99, and Microsoft's FAQ says it starts at 24 GB of unified memory, with configurations of 32, 48, 64 and 128 GB. The 24 GB version ships with Windows 11 Home; the larger ones ship with Windows 11 Pro.

Two pieces of Microsoft's fine print matter for anyone buying this to run models. First, the amount of memory the GPU can address "depends on system configuration and workload and is less than the total." Second, Microsoft describes its FP4 TOPS figures as theoretical and based on the sparsity feature. In other words, the 128GB and 1-petaflop numbers are top-configuration ceilings, not what the starting price buys.

That gap is visible in NVIDIA's own model lineup. Nemotron 3.5 Bolt 74B-A9B 24GB, which NVIDIA released on October 15, uses native 2-bit quantization to fit into 24GB of memory. NVIDIA's launch release says it runs on "supported RTX Spark configurations," not on all of them. A model sized for 24GB is not a comfortable fit on a machine whose total memory is 24GB, so buyers who want local models of that class should look at the larger memory tiers.

The software that ships on day one

NVIDIA pitches RTX Spark as a machine for local AI agents as much as for creative work and games. The release lists:

• Guided installers for agents including GitHub Copilot, Hermes, OpenClaw, Perplexity, Pi and WorkBuddy, without requiring a command line
• NVIDIA OpenShell and Microsoft Execution Containers to control agent permissions
• A PyTorch-CUDA build for Windows on Arm, a native-by-default Python build, a native Conda installer, and support NVIDIA puts at 97% of the top 15,000 Python packages
• NVIDIA Skills, an open-source repository meant to teach agents to use CUDA-X tools and libraries
• NVIDIA PAIR, which lets an RTX Spark machine serve inference to other Windows, Linux and macOS computers on a home network
• Native RTX Spark support in Photoshop and Premiere through beta releases available at launch, with Substance 3D coming later

Microsoft adds a caveat of its own: many apps run natively on Arm, but those that do not run through Prism, Windows' built-in emulator, and buyers who depend on specialized drivers, plug-ins or older utilities should check compatibility before purchase. The remaining 3% of top Python packages, and any compiled extension without an Arm64 build, is where developers are most likely to hit friction.

What we could not confirm

The performance claims in NVIDIA's release, including running frontier models locally, rendering 3D scenes larger than 90GB and playing AAA games at 1440p above 100 frames per second, are NVIDIA's own and come without published methodology in the release. We were not able to read independent benchmarks for this article. Per-model prices other than Microsoft's are not in the sources we reviewed.