NVIDIA DGX Spark

What it is

NVIDIA’s smallest real AI workstation. Grace Blackwell GB10, 20 Arm cores, Blackwell GPU with 6,144 CUDA cores, 128GB of unified LPDDR5X, ConnectX-7 networking, Wi-Fi 7. Ships with the NVIDIA developer stack already on the disk. Two units can be linked.

Who it is for

People who already think in CUDA, want to fine-tune or serve models that do not fit in a 32GB card, and would rather pay once than rent H100 time every week.

What it actually runs

• Comfortable: 20B–70B dense and most 120B-class MoE models, depending on context and quantization.
• Possible: larger models when you pair two Sparks (~256GB).
• Not the win: beating a 5090 on tokens-per-second for anything that already fits in 32GB VRAM. Bandwidth is 273 GB/s.

Standout details

• Paperback footprint: 150 × 150 × 50.5 mm, 1.2 kg.
• Software is the product. DGX OS, CUDA, NGC.
• Memory is soldered.
• Founders price moved from $3,999 at launch (Oct 2025) to $4,699 in Feb 2026 as memory got scarce.

Honest verdict

Buy Spark if CUDA + 128GB in a small cube is the job. Skip it if you want raw decode speed, Windows first, or a machine you can upgrade in two years. Shop GX10 as the same class with a different invoice. Side-by-side: Spark vs GX10.