Mizuho on $AVGO: key points: 1) We see AVGO with the most-capable TH6/Jericho SerDes/Switch for all-toall 1000-1M XPU networking at 200G/lane, SerDes/ASIC/Packaging dominating GOOGL TPU "training" roadmap at 2-3x ASP ~$15K/ASIC, 2) we believe MTK could potentially win the TPU inference chip, despite investors stating MTK is getting the training chip with large die on EMIB [ $INTC ] packaging, and 3) we see the most probable trajectory of AVGO $GOOGL TPUs ramp to 18M+ units in 2028E with AVGO getting ~50% share at ~10-12M units being 3x the current volume (~4.5M today). We see AVGO has the best perf. training SerDes with MTK SerDes limited and needing GOOGL/SMTC (NC) contribution, with potential delays. We believe the v9/v10 will see a mix of both MTK/AVGO to optimize cost and performance, and see AVGO well-positioned into 2028E with upside to 35M TPU, regardless of training/inference chip share. We also see GOOGL purchase commitments (10Q, p55) are expanding significantly to $800 B+ for TPUs.