Electrical Limitations Topple Local AI Rig, Highlighting Industry-Wide DRAM Shortage

A hobbyist building a multi-node local inference rig for AI applications discovered that electrical limitations, not GPU power, became the primary constraint.

The build started with a single RTX 3090 and expanded to include DGX Spark and GB10 machines from Nvidia. However, it was during the 16-RTX-3090 phase that the rig overloaded the household circuit, causing a breaker problem that also blew the house fuses when combined with an electric oven.

The bottleneck was not due to insufficient VRAM or model size but rather the physical limit of the residential electrical system. The Nvidia DGX Spark and GB10 machines use 128GB of unified memory, which can handle models up to about 405 billion parameters when pooled together.

Nvidia's price for the DGX Spark Founders Edition was raised by 18% in February 2026 due to constrained global memory supply. The industry-wide DRAM shortage has pushed up prices on competing AMD Strix Halo-based mini systems as well, with a hardware bill that can run past $90,000 for scaling toward twenty units.