Every Way to Get 512GB of VRAM for Local AI (at Full Context)

512GB of VRAM for local AI doesn't mean 512GB of context. I price every way to get 512GB (Mac Studio M5 Ultra, four NVIDIA DGX Sparks, four RTX PRO 6000s, an EPYC RAM build, 16 AMD MI50s, Strix Halo mini PCs) and work out which models actually run at full context. The number on the box is almost never the number that matters. The model decides. Every token a local LLM reads leaves notes in memory (the KV cache), and that cost per token varies wildly: • DeepSeek V4.1 Flash: 890 bytes per token, so a full 1M-token context is under 1 GB • Muse Spark Open: about 64 KB per token, 69 GB for 1M toke