NEWS → · Hardware · NVIDIA
RTX Spark desktops go on sale: 128GB of unified memory in a mini PC from ASUS, Dell and MSI
Nov 18, 2026
NVIDIA's RTX Spark compact desktops are on sale from today, Nov 18, 2026 — machines from ASUS (the ProArt GR1X pictured), Dell and MSI that put a 20-core Grace CPU and a Blackwell RTX GPU with 6,144 CUDA cores in a mini-PC box, with up to 128GB of unified memory and up to a petaflop of AI performance. NVIDIA says the systems run supported large language models of up to 120 billion parameters locally. Acer, GIGABYTE and HP mini PCs and a Microsoft Surface RTX Spark Dev Box follow later. NVIDIA's announcement lists no per-system prices.

Image: NVIDIA via blogs.nvidia.com
The announcement, from NVIDIA's own blog post of Nov 18, 2026: RTX Spark pairs the Grace CPU and Blackwell GPU on one package with 10GbE networking, aimed at local AI work — agents, fine-tuning and inference — rather than at a datacenter. The 128GB of unified memory is the specification that matters: it is addressable by both the CPU and the GPU, which is what lets a desktop hold a 120B-parameter model that would need a datacenter card or a multi-GPU rental in a conventional PC. NVIDIA also used the post to note native Windows-on-Arm Adobe applications accelerated with TensorRT-RTX and Blender 5.3 support for DLSS 4.5 Ray Reconstruction. Secondary coverage (Gizmodo, Nov 18) puts the desktop ecosystem's starting price at about $4,400 per NVIDIA — that figure is not in NVIDIA's own post and is labelled here as reported, not announced; the earlier RTX Spark laptops, on shelves since October, started at $2,599.
Context from this desk's own coverage: the RTX Spark laptops reached shelves in October alongside a 64GB DGX Spark variant, and NVIDIA showed a four-node DGX Spark HPC Developer Kit at SC26 this week (512GB of combined memory across four GB10 nodes). The desktops extend the same Grace Blackwell package from laptops and developer kits into machines a studio or a small team can simply buy.
What it means if you rent GPUs
For renters this is the buy side of the buy-versus-rent line, priced in one box. A team running an always-on agent or a fine-tuning loop on a single rented card pays by the hour, every hour; a 128GB desktop that holds a 120B model locally pays once. The honest limits are the same ones this desk applies to any single device: one GPU's compute, no NVLink path to a second box for training at scale, and memory that is large but fixed — a model or a context that outgrows 128GB still belongs on rented datacenter cards. What changed today is that the option is actually on sale, from three makers, rather than announced.