DGX Spark ships with Grace Blackwell GB10 superchip offering 128GB unified memory

Nvidia is now shipping DGX Spark, a desktop-sized AI system the company says is built for developers who want to run large models locally rather than only through cloud infrastructure. The system uses Nvidia’s GB10 Grace Blackwell superchip, includes 128GB of unified memory and delivers up to 1 petaFLOP of FP4 AI performance, according to Nvidia product and marketplace pages. Nvidia is marketing the machine for prototyping, fine-tuning and inference on reasoning models, including systems with up to about 200 billion parameters. ### What exactly is DGX Spark shipping now? Nvidia said in a news release that DGX Spark is shipping now after first opening reservations when it introduced the product at GTC in March 2025. The company describes it as a “personal AI supercomputer,” a category it has used for compact systems that package its AI hardware with a preinstalled software stack for local development. (nvidia.com) The current U.S. marketplace listing shows DGX Spark priced at $4,699, while a bundle version with 4TB storage is listed at $9,449. Nvidia’s marketplace page says the system comes with 128GB of unified system memory and is designed to work with AI models locally on the desktop. ### What is inside the GB10 system? The GB10 Grace Blackwell superchip combines Nvidia’s Grace CPU and Blackwell GPU technologies in a single platform aimed at a small-form-factor desktop system, according to Nvidia’s product and hardware documentation. (nvidianews.nvidia.com) Nvidia says DGX Spark is powered by the Grace Blackwell architecture and is intended for developers, researchers and data scientists building and testing large AI models. (marketplace.nvidia.com) Nvidia’s product page says the machine delivers “up to one petaFLOP” of FP4 AI performance and includes 128GB of memory. The company’s marketplace listing adds that the configuration is meant to support local work with models of up to 200 billion parameters, a figure Nvidia presents as a practical ceiling for the box’s target workloads. ### Why does the 128GB unified memory figure matter? (nvidia.com) The 128GB figure is central because Nvidia is pitching DGX Spark as a way to keep larger reasoning models on one local system instead of splitting work across smaller-memory consumer PCs or relying entirely on rented cloud GPUs. Nvidia says the unified memory setup is part of what allows developers to prototype, fine-tune and deploy newer reasoning models directly from the desktop. (nvidia.com) Nvidia’s developer blog said software updates announced at CES 2026 improved performance for both DGX Spark and other GB10-based systems, and pointed to the combination of 128GB unified memory, NVFP4 support and ConnectX-7 networking as features for running larger models locally. That framing matches the company’s current sales language around 200 billion-parameter model support. (nvidia.com) ### Is DGX Spark the only GB10 machine on sale? Acer is already selling a Veriton GN100 AI mini workstation built on the same Nvidia GB10 Grace Blackwell superchip. Acer’s product page says the system includes 128GB of unified memory, up to 4TB of NVMe storage and up to 1 petaFLOP of FP4 AI performance, putting it close to Nvidia’s own published DGX Spark specifications. (developer.nvidia.com) Nvidia’s October 2025 shipping announcement said Acer, Asus, Dell Technologies, Gigabyte, HP, Lenovo and MSI would debut DGX Spark systems. That means Nvidia is treating DGX Spark both as its own product and as a platform design that hardware partners can package into related systems. ### What should buyers watch next? (acer.com) Nvidia’s release notes page for DGX Spark was updated this week, indicating the software stack is still being revised as systems reach users. Nvidia also now hosts a DGX Spark and GB10 user forum, where owners are posting setup questions, performance tuning notes and model benchmarks. The next concrete checkpoints are on Nvidia’s marketplace and documentation pages, where price, storage options, software versions and partner availability are being updated as more GB10 systems reach customers. (nvidianews.nvidia.com) (marketplace.nvidia.com) (docs.nvidia.com)