Budget upgrade for a 31B model? Dual GPUs over PCIe x1 or another approach?

Explain Like I'm Normal

A developer is attempting to run a 31B dense model at 15 tokens per second by daisy-chaining GPUs over extremely limited PCIe x1 lanes. The discussion centers on the trade-offs between VRAM capacity and bus bandwidth bottlenecks for home-grown inference rigs. It highlights the lengths builders go to to avoid cloud costs for privacy-sensitive or high-context tasks.