Overview
- Orders open October 15 through NVIDIA.com and global partner channels, with the first‑party unit priced at $3,999 and OEM variants from Acer, ASUS, Dell, GIGABYTE, HP/HPE, Lenovo and MSI.
- DGX Spark delivers up to 1 petaflop of AI compute with a GB10 Grace Blackwell SoC, 128GB of unified CPU‑GPU memory and up to 4TB NVMe storage, plus NVLink‑C2C and integrated ConnectX‑7 200 Gb/s networking.
- The Ubuntu‑based DGX OS ships preconfigured with NVIDIA’s AI software stack, including CUDA libraries and NIM microservices, enabling out‑of‑the‑box development.
- NVIDIA says the system can run inference on models up to about 200 billion parameters and fine‑tune models up to roughly 70 billion, with the option to link two units for larger 4‑bit inference.
- CEO Jensen Huang hand‑delivered an early unit to Elon Musk at SpaceX’s Starbase as early recipients such as Google, Hugging Face, Meta and Microsoft begin testing, while some commentators question the scale of initial availability.