A tweet from @Techstrongai states that large-scale AI performance no longer depends mainly on GPU count. It claims MoE models and heterogeneous hardware have shifted the bottleneck to the network, which now acts as the performance ceiling. The post adds that infrastructure advantage comes from designing and optimizing the fabric that connects, synchronizes, and feeds accelerators. It links to a Techstrong.ai piece titled "For Massive Models, You Need More Than Just Massive Compute" that discusses how networking determines efficiency as AI clusters scale.