Nvidia unveils its smallest desktop AI supercomputer

Nvidia has begun shipping the DGX Spark, the world’s smallest AI supercomputer for desktop use.

The company says the DGX Spark integrates its Grace Blackwell architecture with GPUs, CPUs, networking, and AI software, aiming to support advanced AI workloads locally.

Major hardware partners, including Acer, ASUS, Dell Technologies, GIGABYTE, HP, Lenovo, and MSI, are releasing their own DGX Spark systems.

Nvidia reports the system delivers up to 1 petaflop of AI performance and 128GB of unified memory, and comes with preinstalled software to support AI model training and inference.

Orders for DGX Spark open October 15, 2025 via Nvidia’s website, with partner systems available globally.

🔗 Source: Nvidia

🧠 Food for thought

Implications, context, and why it matters.

DGX Spark $4,000 price hides memory limits and power draw costs

Marketed at 1 petaflop, benchmarks place DGX Spark about 4x slower than the RTX Pro 6000 Blackwell workstation GPU and behind NVIDIA’s RTX 5090 on large models due to bandwidth limits 1.
Its 273 GB/s memory bandwidth 1 caps production inference throughput, so it fits prototyping and experiments more than full deployment 1.
The compact unit holds stable thermals under load 1. Power comes from external USB-C, with a rated draw near 170W 2, which can complicate office rollout. The desktop power setup is unusual.
Linking two units with ConnectX-7 200 Gigabit Ethernet (GbE) for 405 billion‑parameter models 3 adds gear beyond the $3,999 base 2, so ownership math looks murkier than public cloud GPU options.

Service opportunities around DGX Spark for small labs and healthcare

NYU Global Frontier Lab said DGX Spark’s fit for privacy‑sensitive healthcare work 4, which can drive managed services for procurement, Health Insurance Portability and Accountability Act (HIPAA)–compliant rollout, and ongoing security for medical AI.
NVIDIA’s partners span Dell and HP to Lenovo and ASUS 4, giving it broad channel reach. Integrators can bundle installation, training, plus support for teams without in‑house AI skills.
Support for models up to 70B parameters for fine‑tuning 3 fits schools and smaller biotech firms that want local customization without cloud exposure, creating an underserved opening for turnkey AI lab setup.