“Distributed AI training is highly inefficient if compute nodes are physically separated, with performance dropping by at least an order of magnitude.”