NVIDIA DGX Spark from NVIDIA Corp. - Grace Blackwell GB10 brings local AI into the rack

Reviewed: ad hoc news B2B & Pro desk. Edited and checked on 2026-06-29, 04:59. Details in the imprint.

The NVIDIA DGX Spark sits on a metal rack, fans humming quietly as the Grace Blackwell GB10 superchip chews through a stack of local AI jobs. Status LEDs pulse in a tidy line, and the chassis feels warm but not scorching when you brush its side.

What DGX Spark packs inside

Inside the NVIDIA DGX Spark, the Grace Blackwell GB10 superchip combines CPU and GPU resources with 128 GB of unified memory and fast NVMe storage. One reviewer runs Gigabyte and ASUS variants with 1 TB NVMe drives and reports identical core specs to the reference model.

At the back, the DGX Spark offers four USB-C connectors, a HDMI output and dual ConnectX network adapters that can push up to 200 GB per second. That combination targets small AI factories that want high-throughput networking without filling a full rack.

Power, thermals and daily feel

In practical use, tech reviewer Heavy Metal Cloud says his DGX Spark idles around 25 W and reaches roughly 50 W during heavy RAM and GPU inference loads, with peaks of about 180 W when rendering images and video. Case temperatures peak near 110° Fahrenheit but stay comfortably lower during routine work.

Standing next to the unit, he describes a compact server that runs comparatively quiet for its class, with a steady fan note instead of a harsh roar. That matters if the DGX Spark lives in a small studio or office rather than a sealed server room.

What it can do for AI teams

The DGX Spark targets teams that want local AI capability for language models and generative media without building a full-scale data center. In testing, it serves multiple LLM requests simultaneously and handles new diffusion-based image modes for video and graphics workflows.

Heavy Metal Cloud compares inference speeds to a Mac Studio and finds the DGX Spark about 20 to 50 percent slower in the decoder phase for some models, but notes that the server shines when handling concurrent workloads instead of single interactive sessions.

Where it fits in NVIDIA’s lineup

NVIDIA positions DGX and related systems as part of its broader data center and accelerated computing platform, alongside larger Blackwell-based servers and cloud-scale AI factories. CEO Jensen Huang regularly highlights these systems as the backbone for AI-native companies building their own infrastructure.

For smaller enterprises and advanced creators, DGX Spark sits as an entry point into that ecosystem. It offers a branded, reference-style node that mirrors larger deployments, so teams can prototype locally before scaling up to multi-node clusters.

Pricing and availability

NVIDIA does not push DGX Spark through retail channels, and pricing typically comes via system integrators and enterprise quotes. The Gigabyte and ASUS variants seen in reviews are sold as complete systems rather than bare boards.

For European buyers, availability depends on integrators with NVIDIA partnerships, often bundling DGX Spark with on-site support and networking configuration. That keeps the product in the B2B segment rather than consumer e-commerce.

Investor angle and share listing

Net-net, DGX Spark is one more tile in NVIDIA’s data center mosaic, adding compact local AI servers next to massive cloud deployments. NVIDIA Corp. shares (ISIN US67066G1040) trade on Nasdaq in US dollars, with investors watching data center demand as a key driver.

This article was AI-assisted and editorially reviewed. Product information without guarantee; prices and availability may change at short notice. No investment advice, no buy or sell recommendation. Stock-market transactions involve risks up to total loss.