NVIDIA DGX Spark puts a petaflop on your desk

Back in 2016, Jensen Huang personally delivered the very first NVIDIA DGX-1 to Elon Musk at a tiny startup called OpenAI. That multi-thousand-dollar refrigerator-sized box helped ignite modern generative AI.

This week, history repeated itself with a twist. Huang showed up at SpaceX’s Starbase in Texas to hand-deliver another machine to Musk, who was busy handing out donuts to kids and grabbing pizza in the cafeteria. Only this time, the machine didn't require two people to carry it. It weighs just 1.2 kg and is about the size of a piece of origami paper.

Meet the DGX Spark, which NVIDIA is calling the world’s smallest AI supercomputer. It starts shipping on October 15 directly from NVIDIA and through partners like Dell, HP, Lenovo, ASUS, and Acer.

A petaflop on your desk

Despite being the thickness of a hardcover book, the DGX Spark packs serious enterprise iron.

At the center sits the NVIDIA GB10 Grace Blackwell Superchip, pumping out up to 1 petaflop of AI performance at FP4 precision. It pairs that compute with 128GB of coherent CPU-GPU unified memory, tied together by NVLink-C2C to deliver five times the bandwidth of fifth-generation PCIe.

In plain English: you can now run local inference on AI models with up to 200 billion parameters right on your desk. You can also fine-tune 70-billion-parameter models without paying hourly cloud rental fees or uploading proprietary data.

It even features ConnectX-7 200 Gb/s networking. Dell demonstrated that hooking two of its GB10 systems together creates a single workstation capable of running 400B models.

Out-of-the-box developer stack

Hardware is only half the battle, and NVIDIA is shipping the system fully loaded with its CUDA stack and NIM microservices.

Developers can immediately run locally optimized workloads, such as fine-tuning Black Forest Labs' FLUX.1 models, orchestrating vision agents with NVIDIA Cosmos, or deploying Qwen3 chatbots.

Nearly every major local AI tool is already optimizing for the box, including Ollama, LM Studio, Hugging Face, ComfyUI, Docker, and JetBrains. NYU's Global AI Frontier Lab noted that having desktop petaflop computing is transformative for prototyping in privacy-sensitive sectors like healthcare.

Beyond online orders, systems will also be sold through retail shelves at Micro Center stores in the US.

Why it matters

We think DGX Spark is NVIDIA’s smartest hardware flex in years.

For developers, local AI has long hit a hard memory ceiling. Running anything north of 30B parameters smoothly usually meant stringing together multiple bulky consumer GPUs or surrendering to cloud bills. DGX Spark’s 128GB unified pool solves the memory wall locally, making it an immediate temptation for AI researchers, robotics engineers, and privacy-conscious shops who want rapid prototyping without latency.

For the broader industry, NVIDIA is locking down the developer pipeline from the ground up. By marrying proprietary Grace Blackwell silicon with NVLink-C2C in a miniature desktop box, it pushes traditional x86 workstation setups into an awkward corner.

The catch? That headline 1 petaflop figure leans heavily on FP4 precision, which won't fit every demanding training or fine-tuning workflow without quantization trade-offs. Thermal performance inside a tiny 1.2 kg chassis will also be put to the test under heavy loads. But if the pricing lands reasonably, NVIDIA may have just put a miniature supercomputing node on every serious developer’s desk.