NVIDIA’s next big bet on accessible AI is almost here. In a high-profile handoff that signals just how important this product is, CEO Jensen Huang personally delivered a DGX Spark to Elon Musk, echoing the moment years ago when one of the first DGX-1 systems was handed over during Musk’s OpenAI era. The visit reportedly coincided with activity around the 11th Starship test and included a casual cafeteria chat where Huang reflected on that earlier milestone and how Spark pushes the mission further.
Unveiled at CES 2025, the DGX Spark is a compact mini‑supercomputer designed to put serious AI horsepower within reach of developers, researchers, startups, and advanced creators. Retail availability is now slated for October 15, following a brief delay from the original July window due to the complexities of its custom GB10 system-on-chip, co-developed by NVIDIA and MediaTek. When it lands, you’ll be able to order it directly as well as through major vendors including Acer, ASUS, Dell Technologies, GIGABYTE, HP, Lenovo, and MSI.
Key DGX Spark highlights:
– GB10 Grace Blackwell Superchip delivering up to 1 petaflop of AI performance at FP4 precision
– 128GB of unified CPU-GPU memory so you can prototype, fine-tune, and run inference locally without switching machines or cloud instances
– NVIDIA ConnectX networking for clustering and NVLink‑C2C providing up to 5x the bandwidth of PCIe
– NVMe storage for fast data access and HDMI output for visuals
The pitch is straightforward: workstation-sized AI performance without relying on the cloud. That unified memory pool is especially compelling for developers who want to iterate quickly on fine-tuning and inference, while NVLink‑C2C and ConnectX open the door to scaling out multiple units for bigger projects. It’s a serious tool with a serious price tag—around $3,999—but it’s aimed at professionals who need reliable, on‑prem AI compute in a compact footprint.
With the DGX Spark heading to retail by mid‑October, the countdown is on. If you’ve been waiting for a small-form-factor system built for modern AI workloads—one that blends portability, speed, and scalability—this looks like a standout option in 2025’s AI hardware landscape.
