Memory is king • Elon focused on HBM for terafab as the most important bottleneck(not NAND or HDDs) • Gavin Baker: HBM could be 3–40% of hyperscaler capex in 2027 • Jukan: memory may matter more than Nvidia going forward • Micron CEO: robotics may require 10x more memory • Memory is still cyclical, but HBM is becoming less commodity-like Main longs • SOXL • HBM: SK Hynix, Samsung, Micron Barbell “Anything that helps soften the memory blow should do well over the next year” Would approach these more technically • MRVL: custom ASIC + HBM architecture for hyperscalers • PENG: CXL / KV-cache servers to offload GPU HBM pressure • QCOM: HBC, near-memory compute, possible HBM alternative + robotics angle • AMD: MI350/MI450 = non-NVDA HBM capacity monster