NVIDIA DGX Spark: The Home AI Server That Fits Under a Monitor
128 GB unified memory. Blackwell. On your desk.. NVIDIA’s personal DGX is the first home-sized box that treats 70B-class models as a Tuesday, not a science project.
A home AI server used to mean a used Xeon, three 3090s, and a marriage counselor. The DGX Spark is NVIDIA’s attempt to replace that hobby with an appliance: GB10 superchip, 128 GB of unified LPDDR5X, a Blackwell GPU, and a 1.2 kg chassis that sits under a display like a Mac Studio that went to grad school.
The useful numbers: enough unified memory to hold real models (Llama-class 70B in 4-bit is not a stunt), NIM and CUDA as first-class citizens, Ollama a command away, and 200 GbE on the serious SKUs if you are the person who already has a 100-gig rack at home. Reviews put the street price in the mid-four-thousands. That is not impulse-bin money. It is “I am done renting tokens” money.
What it is not: a silent living-room gadget, a Windows gaming box, or a magic ChatGPT replacement on day one. It is a Linux workstation that happens to be the most concentrated local-inference machine you can buy without filling a closet. Pair it with Open WebUI and it becomes the house model. Pair it with nothing and it is an expensive paperweight that happens to be gorgeous.
InteliToys coverage of home AI servers starts here because this is the object that made the category mainstream. Framework Desktop and Mac Studio are the other two corners of the triangle. Read all three before you spend.
InteliToys is editorial. We do not sell NVIDIA DGX Spark. Watch the source tape in the branded lightbox below, then argue about it in Gear Junkies.