Shard claims a 744B GLM-5.2 run split across six RTX PRO 6000 nodes in different US states, with a coordinator using a small CUDA-graphed draft model and receipts for GPU UUIDs, IPs, RTTs, token hash, and lossless checks. The security hook is practical rather than dramatic, because the repo admits encrypted pickle-free transport is only part of privacy and participating nodes still see activations.