NVIDIA has launched the DGX Spark, the world’s smallest AI supercomputer, to bring petaflop-level performance to developers.
The compact desktop enables local AI development by supporting inference on models with up to 200 billion parameters and fine-tuning of models up to 70 billion parameters, reducing reliance on cloud infrastructure.
In a symbolical gesture, NVIDIA CEO Jensen Huang personally delivered the first unit to Elon Musk at SpaceX’s Starbase facility in Texas. This act echoes Huang’s 2016 delivery of the original DGX-1 to Musk’s then-startup, OpenAI, which has played a foundational role in the development of ChatGPT and the broader AI revolution.
“DGX-1 launched the era of AI supercomputers and unlocked the scaling laws that drive modern AI. 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,” said Huang.
Based on the NVIDIA Grace Blackwell architecture, DGX Spark integrates GPUs, CPUs, networking, CUDA libraries, and the full NVIDIA AI software stack into a single desktop unit to make it a self-contained AI development platform.
With up to one petaflop of AI performance and 128GB of unified CPU-GPU coherent memory, the system delivers high-speed processing through NVIDIA ConnectX-7 200 Gb/s networking and NVLink-C2C technology, which provides five times the bandwidth of PCIe Gen5.
The pre-installed NVIDIA AI software stack lets developers start projects immediately, providing access to tools such as models, libraries and NIM microservices for local workflows. These include customising Black Forest Labs’ FLUX.1 models for image generation, building vision search agents using the NVIDIA Cosmos Reason model, or creating optimised AI chatbots with Qwen3.
The system is particularly valuable for privacy- and security-sensitive applications, such as healthcare research, where data must remain on-premise.
DGX Spark is targeted at AI developers, researchers and enterprises seeking to accelerate agentic and physical AI development on a local machine. It is ideal for startups, academic labs and enterprise teams that require high-performance computing but lack access to large-scale data centres.
Technology players such as Hugging Face, Docker, Google, Microsoft, Meta, and Roboflow are among the early recipients testing and optimising their tools for the platform.
Starting October 15, 2025, DGX Spark will be available for order directly from NVIDIA.com and through global channel partners. Partner systems will be offered by Acer, ASUS, Dell Technologies, GIGABYTE, HP, Lenovo, and MSI.
Entelechy Asia 