Unboxing the NVIDIA DGX Spark: An AI Supercomputer on My Desk

The box arrived yesterday. It’s smaller than I expected — a square package that sits comfortably on the corner of my desk, next to my keyboard. Inside was the DGX Spark, and a few accessories.

Not a lot. Because the magic isn’t in the accessories. It’s in what’s inside that white, minimalist case.

What Is the DGX Spark?

NVIDIA calls it “an AI supercomputer on your desk.” That’s not entirely marketing hyperbole. This is a purpose-built desktop AI workstation powered by the NVIDIA GB10 Grace Blackwell Superchip — a 20-core ARM processor with a custom NVIDIA GPU, all sitting on a single chip.

The idea is simple: run real AI workloads locally, without needing a cloud GPU, a rack of servers, or a data centre. For AI development, inference, fine-tuning, edge computing — it’s all there. On your desk.

What’s in the Box?

The unboxing is surprisingly sparse. There’s the DGX Spark unit itself — a clean, white, square slab measuring 150mm x 150mm x 50.5mm. It’s small enough that you’d think it’s a toy, until you read the specs.

Inside the box:

• The DGX Spark unit (GB10 superchip)
• Power supply and cable
• Quick start guide

That’s it. No cables for peripherals (USB-C, Ethernet, HDMI/DisplayPort are built into the unit). No extra storage modules. No fan. Because it doesn’t need one — the GB10 is designed for passive thermal management.

It looks like a high-end tablet. It runs like a workstation.

The Specs That Actually Matter

Forget the numbers for a second — here’s what this thing means.

You know how AI keeps getting bigger but your laptop keeps getting left behind? This device is the “finally” answer to that. Plug it in, load a model, and run inference locally — your own data, your own machine, zero cloud dependency. No API calls leaving your house. No waiting for a remote server to process your prompt. It just works. On your desk.

128 GB of unified memory. Yes, really. A device this size with 128 gigabytes shared between CPU and GPU. It means 200B parameter models load and run on your desk! Models that a few years ago needed a rack of GPUs and a dedicated data centre room — now they hum quietly next to your keyboard.

1 petaFLOP of AI compute. A top-tier consumer GPU gives you around 1 TeraFLOP of AI performance. So you’re sitting on a desktop AI machine with roughly the same firepower as one of those, except this thing is the size of a tablet and draws no more power than a regular PC.

The catch isn’t the hardware — it’s the ecosystem. It’s ARM-only, so some x86-optimised software needs workarounds. It draws a fraction of the power of an RTX 4090 while running AI workloads on the same hardware footprint. And it’s not a general-purpose workstation. But for anyone who wants to seriously play with AI locally — that gap just disappeared.

Still, for what it’s designed to do, it’s genuinely impressive. A pocket-sized AI supercomputer that can run models that previously required data-centre GPUs. That’s the kind of hardware shift that takes years to appreciate and minutes to understand once you see it in person.

Final Thoughts

Unboxing it was a reminder of how far things have come. Before the Spark, running a 70B parameter model locally was science fiction. This easily changes today: there’s a device you can plug into a wall socket and start fine-tuning models before your coffee gets cold.

The unboxing photo above is the DGX Spark — a small, square, white slab that punches way above its weight class. It doesn’t look like much. But look closer, and it’s the most interesting piece of hardware I’ve seen this year.