NVIDIA CEO Jensen Huang hand-delivered the company's newly launched DGX Spark AI supercomputer to Elon Musk at SpaceX's Starbase facility in Texas on Monday. The delivery marks the commercial rollout of what NVIDIA characterises as the world's smallest AI supercomputer, designed to bring petaflop-class performance to developers, researchers and creators outside traditional data center environments.
DGX Spark delivers one petaflop of AI performance at FP4 precision within a 1.2-kilogram chassis. The system features NVIDIA's GB10 Grace Blackwell Superchip with 128GB of unified CPU-GPU memory, enabling users to prototype, fine-tune and run inference on models with up to 200 billion parameters locally. The hardware includes NVIDIA ConnectX networking for clustering capabilities, NVIDIA NVLink-C2C providing five times PCIe bandwidth, NVMe storage and HDMI output.
The handoff at SpaceX occurred as the company prepared for the 11th test of Starship. Huang referenced NVIDIA's delivery nine years ago of the first DGX-1 system to OpenAI, positioning DGX Spark as an evolution of that enterprise AI strategy. "Imagine delivering the smallest supercomputer next to the biggest rocket," Huang said. The system ships with NVIDIA's complete AI software stack, including frameworks, libraries, pretrained models and NVIDIA NIM microservices supporting workflows such as customizing FLUX.1 image-generation models, building vision search agents with NVIDIA Cosmos, and deploying Qwen3 chatbots.
Seven major PC manufacturers—Acer, ASUS, Dell Technologies, GIGABYTE, HP, Lenovo and MSI—will distribute DGX Spark systems. Initial deployments include Ollama in Palo Alto for local large language model operations, NYU Global Frontier Lab for privacy-sensitive algorithm development, Zipline for autonomous delivery applications, Arizona State University for robotics simulations, and artist Refik Anadol's studio for AI-integrated creative work.
DGX Spark addresses enterprise requirements for on-premises AI development by eliminating dependency on cloud instances for model prototyping and inference workloads. The unified memory architecture and local processing capability support organisations managing data privacy constraints while maintaining access to frontier model performance. The multi-vendor distribution strategy through established PC OEMs provides enterprise procurement channels and support infrastructure.
DGX Spark represents NVIDIA's strategic expansion from data centre GPU dominance into edge AI and developer workstation markets. The system's form factor and performance specifications target organisations seeking to decentralise AI compute resources while maintaining model performance parity with data center systems. The 128GB unified memory and 200-billion parameter local inference capability position the platform for enterprise use cases requiring data sovereignty compliance. General availability through NVIDIA.com and global partners begins October 15, 2025, establishing immediate market access across research institutions, ISVs and creative production environments.