NVIDIA is putting supercomputer-class AI performance onto the desks of developers worldwide with the arrival of DGX Spark, a compact system designed to deliver the full power of the company’s AI ecosystem in desktop form. Built on the Grace Blackwell architecture, DGX Spark combines NVIDIA GPUs, CPUs, networking, CUDA libraries, and preinstalled NVIDIA AI software into a unified platform ready for local AI development and experimentation.
With up to one petaflop of AI performance and 128 GB of coherent CPU–GPU memory, DGX Spark makes it possible to run inference on models approaching 200 billion parameters and fine-tune models up to 70 billion parameters — directly on the desktop. NVIDIA positions the system as a bridge between the limits of personal workstations and the scale of cloud data centers, offering researchers and developers the ability to run advanced AI agents and physical AI workloads locally.
In a symbolic moment, Jensen Huang personally delivered the first DGX Spark to Elon Musk at SpaceX’s Starbase in Texas — a nod to 2016, when Huang handed Musk one of the first DGX-1 units at OpenAI, a machine that helped catalyze the development of ChatGPT. “With DGX Spark, we return to that mission — placing an AI computer in the hands of every developer to ignite the next wave of breakthroughs,” Huang said, as quoted on NVIDIA’s newsroom.
Major hardware partners including Acer, ASUS, Dell Technologies, GIGABYTE, HP, Lenovo, and MSI are debuting DGX Spark systems, while software companies such as Anaconda, Docker, Hugging Face, Meta, Microsoft, and JetBrains are optimizing their tools for the new platform. Researchers at the NYU Global Frontier Lab described the desktop supercomputer as enabling “peta-scale computing on our desktop,” accelerating experimentation in sensitive domains such as healthcare.
DGX Spark integrates GB10 Grace Blackwell Superchips, ConnectX-7 200 Gb/s networking, and NVLink-C2C interconnects — offering five times the bandwidth of PCIe Gen 5 — while providing seamless access to NVIDIA’s software ecosystem, including NIM microservices and prebuilt models like Cosmos Reason and Qwen3. Developers can start working out of the box, customizing generative and multimodal models without needing to rely on cloud compute.
Orders for DGX Spark open October 15 on NVIDIA.com, with partner systems available globally through vendors and retail channels such as Micro Center. For developers who have long been constrained by limited local resources, DGX Spark represents a new class of desktop computing — one designed to bring large-scale AI experimentation closer to where innovation happens: right at the developer’s desk.