Katarzyna Petru
NVIDIA has just begun global shipping of the DGX Spark – the smallest AI supercomputer in the company's history. The system aims to give artificial intelligence creators server-class local computing power packed into a desktop format. In practice, this means 1 PFLOP of computing power and 128 GB of unified memory, allowing them to run models with up to 200 billion parameters without the need to use the cloud.
The launch wasn't without its symbolism – Jensen Huang, CEO of NVIDIA, personally delivered the first unit to Elon Musk at SpaceX, referencing 2016 when the same duo collected the first DGX-1 system, from which ChatGPT later emerged.
“In 2016, the DGX-1 gave scientists their own supercomputer. Now, the DGX Spark is set to put AI in the hands of every developer” – Jensen Huang said.
AI power in a desktop format
The new DGX Spark is built on the Grace Blackwell architecture and equipped with the GB10 Superchip, which combines GPU and CPU with a bandwidth 5x greater than PCIe 5.0. The entire system consumes just 240 W of power, weighs 1.2 kg, yet offers petascale data processing.
Included is a complete environment of NVIDIA AI Stack – CUDA, NVLink-C2C, ConnectX-7 (200 Gb/s), libraries, models, and microservices NVIDIA NIM, allowing for immediate AI project deployment right out of the box.
The first units are going to partners such as Acer, ASUS, Dell, Lenovo, MSI, HP, GIGABYTE, as well as to Anaconda, Meta, Microsoft, Hugging Face, and JetBrains. NVIDIA announces that DGX Spark will become the new standard for creating AI agents and physical models, also in local applications.
Comparison: from DGX-1 to DGX Spark
Specification | DGX-1 (2016) | DGX Spark (2025) |
|---|---|---|
GPU Architecture | NVIDIA Pascal | NVIDIA Blackwell |
GPU Memory | 128 GB (16 GB per GPU) | 128 GB unified system memory |
AI Performance | 170 TFLOPS (FP16) | 1 PFLOP (FP4) |
Power Consumption | 3200 W | 240 W |
Dimensions | 866 × 444 × 131 mm | 150 × 150 × 50.5 mm |
Weight | 60.8 kg | 1.2 kg |
Price | 129,000 USD | 3,999 USD |
New era for AI creators
Thanks to Spark, developers can locally train and fine-tune models with up to 70 billion parameters, create AI agents, tools for computer vision, or personal chatbots optimised for Grace Blackwell. The system is fully ready to work with models such as FLUX.1, Qwen3, or Cosmos Reason.
As Professor Kyunghyun Cho from NYU Global AI Frontier Lab put it:
“DGX Spark gives us petascale power on the desk. It’s a completely new way of conducting AI research — even in sensitive areas like medicine.”
From 15th October, DGX Spark can be ordered from NVIDIA.com, and partners such as Acer, ASUS, Dell, HP, Lenovo, and MSI are offering their own versions of this computer. It will also be available in the Micro Center network in the USA, and global distribution will start in the coming weeks.
