DGX Spark stirs custom-build debate

DGX Spark is best understood as a local AI appliance, not a generic tinkerer box. NVIDIA has optimized the common AI workflows, but the tradeoff is living inside a fixed Arm/Ubuntu/NVIDIA ecosystem.

• –NVIDIA's own guides frame ComfyUI as a 45-minute setup and the FLUX LoRA flow as a 1-hour Docker workflow, so it is usable day-to-day but still a real Linux box, not a plug-and-play appliance.
• –Third-party trainers like kohya or AI-Toolkit are where ARM64 package and Docker quirks are more likely to show up, so the official NVIDIA stack will usually feel smoother.
• –The fixed 128GB unified memory is the killer feature, but also the lock-in: you are buying capacity you cannot expand later, so future-proofing depends on buying enough headroom on day one.
• –NVIDIA's playbooks also cover JAX, vLLM, SGLang, RAG, and data-science tasks, so Spark is more than a one-trick inference box if your work stays in the AI/Python world.
• –For non-technical bosses, the clean pitch is "supported local AI appliance vs DIY workstation": less setup and cloud spend now, less flexibility and upgrade room later.