NVIDIA DGX Spark: an AI supercomputer on your desk
A machine the size of a Mac Mini that runs 200-billion-parameter models locally. What the DGX Spark really changes for an SME — minus the marketing.
By Nacim Moudjeb6 min3
An AI supercomputer in a 1.2 kg box
In October 2025, NVIDIA put on sale a machine that would have looked absurd two years earlier: a computer the size of a Mac Mini (150 × 150 × 50 mm, 1.2 kg) able to run AI models that until then only datacenter servers could handle. The DGX Spark. And behind the buzz, there's a real shift for anyone who wants to do AI without sending everything to the cloud.
Let's look at what it actually is, and above all who it's for.
What's in the box
The figures come from NVIDIA's official spec sheet:
GB10 Grace Blackwell chip: a 20-core Arm CPU and a Blackwell GPU on the same chip.
128 GB of unified memory (LPDDR5x), shared between CPU and GPU — that's what lets it load big models.
Up to 1 PFLOP (a quadrillion operations/second) at FP4 precision.
4 TB encrypted SSD, 240 W at the wall (it plugs into a normal socket).
ConnectX-7 networking at 200 Gbit/s, which lets you link two units.
NVIDIA DGX OS (Linux-based).
In practice: it does inference up to 200 billion parameters, fine-tuning up to 70 billion, and by linking two machines you reach 405-billion models. Models like Llama, Qwen, Gemma or DeepSeek run on it out of the box.
The price, which moved
Announced at $2,999 at CES in January 2025, it shipped at $3,999 on October 15, 2025, then climbed to $4,699 on the NVIDIA Marketplace in early 2026. It's also built by ASUS, Dell, HP, Lenovo and Acer. So budget around $4,500-5,000 for the current version — the price of a good workstation, not a server.
What it really changes (and what it doesn't)
Let's be honest, since the marketing isn't: the DGX Spark is not a production server for serving thousands of users. Its memory bandwidth (273 GB/s) is modest next to a datacenter GPU like an H100 (more than ten times higher). The result: on a very large model, it generates text slowly. This is not the machine that will run your public chatbot at full load.
What it changes is something else, and it matters for an SME:
Prototyping AI locally, on a desk, without renting cloud GPU by the hour. You test, iterate, break things, with no climbing bill.
Keeping data in-house. For sectors where nothing should leave (healthcare, legal, sensitive data), that's a direct win — the subject of our piece on self-hosting.
Fine-tuning models up to 70B parameters without a datacenter. For the cases where fine-tuning is genuinely warranted, it's a serious shortcut.
Should you buy one?
If you're an SME curious about AI, the DGX Spark isn't a reflex purchase. It's a development and privacy tool, not a magic wand. The right move: first know which problem you want to solve, then decide whether local — and so a machine like this — is the right answer.
That's exactly what we do upfront: we look at your case, your data, your constraints, and tell you honestly whether a DGX Spark makes sense for you or a good API is enough. Let's talk.