NVIDIA DGX Spark, GB10 Grace Blackwell Desktop AI System

A compact desktop AI development system, a small, power-efficient personal computer for AI development, not a rack server despite the DGX name. It is built on the NVIDIA GB10 Grace Blackwell Superchip with 128GB of unified memory and 1,000 AI TOPS, and it sits on a desk and plugs into a standard outlet. It lets an individual developer run, prototype, and fine-tune large AI models locally.

No, and that is the key thing to understand. Despite the DGX name and "supercomputer" framing, the DGX Spark is a personal desktop development machine, not a rack-mounted server or a data center system. The data center DGX systems (DGX H100, H200, B200) draw kilowatts in a data center rack and train at scale; the DGX Spark sits on a desk, draws a fraction of the power, and is for local development. It is the desk-side counterpart to those systems, not one of them.

Large AI models, locally. With 128GB of unified memory and 1,000 AI TOPS, it can prototype, fine-tune, and run inference on current large models, including reasoning models from developers like DeepSeek, Meta, and Google, with support for up to 200 billion parameters on the device. That is well beyond what a normal workstation handles, which is what makes it a serious local AI development tool.

Through a shared software stack and workflow. The DGX Spark ships with the NVIDIA AI software stack pre-installed, the same stack that runs on the data center DGX systems, so a developer prototypes and fine-tunes locally on the Spark, then hands the work off to a data center or cloud DGX system for full-scale training or production deployment, without re-engineering the software. Build on the Spark, scale on the data center systems.

Yes. NVIDIA previewed this system under the name Project DIGITS before launching it as the DGX Spark. If you followed the early announcements under that name, the DGX Spark is the shipping product, the compact GB10 Grace Blackwell desktop AI development system.

The NVIDIA GB10 Grace Blackwell Superchip (Grace CPU plus Blackwell GPU), 128GB of unified memory, 1,000 AI TOPS of AI performance, support for models up to 200 billion parameters, and the NVIDIA AI software stack pre-installed, in a compact desktop form. MillionMiner confirms the full configuration, including storage, connectivity, and exact power, in the quote.

AI engineers, researchers, and developers who want serious local AI capability, to prototype, fine-tune, experiment with, and run inference on large models at their desk, rather than queuing for shared data center resources or sending everything to the cloud. It suits individuals and teams building AI applications who value local development with a clean path to scale.

Control, iteration speed, and cost. Local development on the Spark means no per-hour cloud charges for prototyping, no queueing for shared resources, full control over data and environment, and fast iteration at the desk. When the work is ready to scale, the pre-installed NVIDIA software stack makes the handoff to data center or cloud straightforward. For the development and fine-tuning stages, local is often faster and more economical; for full-scale training, the data center is the tool.

Yes. As a current NVIDIA product, the DGX Spark ships new. MillionMiner confirms the configuration in the quote and tests every unit before shipment.

Submit your needs through the quote form. A MillionMiner specialist confirms the configuration and quantity, with delivery. Every unit is tested before shipment and shipped worldwide DDP with duties handled, and the specialist can advise on the data center DGX systems and hardware the Spark hands off to if you also need full-scale training or production infrastructure.