NVIDIA DGX Spark 64GB Gives Developers More Ways to Build and Scale Local AI

NVIDIA says a 64GB DGX Spark configuration ships from Acer, ASUS, Dell, Gigabyte, HP, and MSI starting Friday, October 23, with prices beginning at $4,999. Same GB10 Grace Blackwell Superchip, DGX OS, and NVIDIA AI stack as the 128GB model; the pitch is on-device agents up to roughly 100-billion-parameter models without a cloud dependency for every task.

Two units can cluster over the built-in ConnectX-7 NIC with NVIDIA Sync Cluster Assistant, pooling memory to 128GB and, in NVIDIA’s Qwen 3.8 27B test, delivering up to 1.7x performance versus one box. End-of-month Model Launcher promises one-click local launch across a cluster.

We’re for this class of product when it makes private, on-desk inference boring and reliable. The interesting move isn’t another GPU SKU — it’s treating clustering as a consumer-adjacent workflow: plug two boxes, let Sync configure the fabric, keep the same software stack. That’s useful AI for builders who want agents running overnight on their own data.

The hedge is economic and cultural. Five thousand dollars is “accessible” only in the workstation sense, and partner exclusivity means NVIDIA is still a platform tax on the whole chain. Local AI that only works inside NVIDIA’s OS and agent toolkit is freedom with a logo on the door.

I’m watching whether two-node Sync becomes something people actually use, or a demo that collapses into single-box hobbyist kits. If Blender and the agent playbooks land cleanly, this is a real alternative to renting tokens for every experiment. If not, it’s a nicer Mini PC with a press kit.

DGX Spark sits in NVIDIA’s local-AI push alongside RTX Spark Windows PCs from major OEMs later this month. The company is also pointing creators at open models like Alibaba’s Qwen-Image-2.1 on RTX, Spark, and Station hardware.