DGX Spark clones draw homelab buyers

This is less news than a useful signal: local AI hardware demand is moving from hobbyist GPU towers toward compact unified-memory workstations, but the price-to-performance math is still messy.

• –NVIDIA DGX Spark and GB10 partner boxes target developers who want CUDA, 128GB unified memory, and an appliance-like local AI stack
• –ASUS Ascent GX10 and Dell Pro Max with GB10 are the closest Spark-style alternatives; HP Z2 Mini G1a is more of an AMD unified-memory rival than a true clone
• –For local LLM inference, memory capacity often matters more than raw GPU speed, especially around 70B-120B quantized models
• –For learning training and fine-tuning, a used multi-GPU desktop can still beat mini workstations on throughput, serviceability, and cost
• –The $3,500 budget sits in an awkward gap: enough for serious local inference, but below many current DGX Spark-class retail prices