Crypto rigs pivot to local LLM infrastructure

Repurposing mining rigs is the ultimate "cheap VRAM" hack, but it’s a race against energy bills and bandwidth bottlenecks.

• –VRAM is the primary constraint for LLMs, making 8GB legacy cards far more valuable than faster 4GB modern cards for large model loading
• –Mining rigs often use PCIe x1 risers, which cripples tensor parallelism; users must rely on pipeline parallelism to avoid massive performance penalties
• –Software compatibility is the "hidden tax," requiring legacy ROCm versions or Vulkan backends to support aging AMD Polaris architecture
• –While 64GB of VRAM for under $1000 sounds like a bargain, the power draw and heat of 8+ vintage GPUs may cost more over time than a single used RTX 3090