Interview: Microsoft's Maia GM on Maia 300, Taiwan and HBM supply

Microsoft pinning Maia 300 as the primary inference engine for Copilot is a structural decoupling of training and inference compute stacks. Inference demands far less HBM bandwidth than training but is far more sensitive to latency and energy-per-token, so Maia 300 almost certainly favors large on-die caches and dedicated data paths over raw memory capacity. The ripple hits HBM suppliers first: inference chips consume roughly 40-60% of the HBM per die versus training accelerators, yet shipment volumes are multiples higher. Net demand doesn't shrink; it redistributes from few high-capacity packages to many mid-capacity ones, reshaping SK Hynix and Samsung HBM3E order books. On manufacturing, CoWoS advanced-packaging capacity in Taiwan, China remains the binding constraint. Microsoft retaining Nvidia and AMD GPUs is not hedging; it is an admission that Maia cannot replicate CUDA's operator library for training within this cycle. Nvidia's real exposure is the structural erosion of inference-card revenue once Microsoft internalizes inference costs. Over the next 12-24 months, the HBM4 production window in H1 2026 is the generational inflection point. If Maia 300 integrates HBM4 first, it gains architectural leverage over Nvidia's B300. If SK Hynix differentiates supply allocation between the two, the bargaining-power map of the entire AI memory chain gets redrawn.