DGX Spark

DGX Spark is a computer from the company Nvidia, about the size of a thick book. It is not meant for gaming or office work, but for developing artificial intelligence. Such programs need an extreme number of computing steps simultaneously, and that is exactly what the hardware in this device is designed for. Nvidia first introduced it in early 2025 under the name Project DIGITS, and from October 2025 it was available for purchase under the name DGX Spark. The price is around 4,000 US dollars, far above that of a normal laptop. The device is especially interesting for people who want to build and test AI software without having to rent computing time online every time.

A data center on the desk

Modern AI models are huge collections of numbers, often many billions of them. These numbers have to fit completely into the computer’s memory, otherwise the model won’t run at all. Normal graphics cards for private users usually have 8 to 24 gigabytes of memory for this. That is only enough for smaller models. DGX Spark offers 128 gigabytes, which the processor and graphics unit share.

This shifts who can actually work with large models at all. Until now, developers had to rent servers from providers like Amazon or Microsoft for this. That costs money continuously, and the data leaves the company’s own premises. For firms with sensitive data, for example in medicine, this is a real problem. A device in one’s own office resolves this conflict.

For Nvidia itself, the product is also strategically important. The company earns its money almost entirely from AI hardware for large data centers. With DGX Spark, it is trying to establish its technology among individual developers and small teams as well. Whoever develops on this device will later almost automatically continue working with Nvidia servers.

What’s inside the case

The core of the device is a chip called GB10 Grace Blackwell. It contains two parts: a normal main processor with 20 computing cores and a graphics unit. The graphics unit is the actual engine, since it can carry out thousands of simple calculations in parallel. Exactly these kinds of calculations are what an AI model needs when it generates an answer word by word.

The memory in DGX Spark is organized unusually. The processor and graphics unit access the same memory block, which is called unified memory. In a normal PC, by contrast, data first has to be copied back and forth between main memory and the graphics card. This copying costs time and limits model size. If it is eliminated, significantly larger models fit into the device.

A comparison helps put this in perspective. A high-end gaming graphics card computes faster than DGX Spark for small models. The Spark doesn’t win through speed, but through capacity. It’s more of a big truck than a sports car. And if you couple two devices via a fast network connection, even larger models can be loaded.

Who actually uses a device like this

In practice, such computers are found in research groups at universities, in software start-ups, and in the AI departments of larger companies. A typical task is fine-tuning: you take a finished, publicly available model and retrain it with your own data. This turns a general-purpose language model, for example, into an assistant for legal texts.

In news reports, you’ll usually encounter DGX Spark in connection with Nvidia’s business figures or with the question of whether AI computing power is becoming more decentralized. Star power also plays a role: Nvidia CEO Jensen Huang personally delivered the first devices, among others to Elon Musk and to OpenAI. Such images are marketing, but they show how important the product is for the company.

A common misconception is the idea that you could train a model like ChatGPT from scratch on a device like this. That is out of the question. For that you need thousands of such chips over weeks on end. DGX Spark is meant for adapting and experimenting, not for the initial large-scale training.