TPU

Summary * DeepSeek, Huawei, and then AVGO in 2025 pulled forward the timeline for optical adoption, flipping the “copper-first” consensus and reshaping cluster design priorities. * DeepSeek’s P/D MoE inference shows scale-out fabrics can handle high-performance inference across hundreds of GPUs, reducing the marginal value of ever-larger NVLink scale-up nodes. * Huawei’s optics-first CM384 “superpod” pushes scale-up-like behavior across racks, narrowing the gap with NVDA despite weaker chips