[FLASH] Microsoft unveils Maia 300 on TSMC N2P, plans multi-gigawatt Maia deployments from 2027
Microsoft's official blog (Scott Guthrie, Oct 12) says the third-generation Maia 300 has eight HBM4 stacks and runs at 950 W. Production deployments begin in 2027, and Maia is to become the 'primary AI inference system' for Microsoft products. Microsoft projects about 40% more tokens per megawatt than 'leading GPU-based systems'; that figure is a projection against unnamed rivals. DigiTimes reported the TSMC N2P detail independently. In its own interview, Seoul Economic Daily quoted Maia GM Andrew Wall saying HBM supply is tightly constrained and Microsoft wants to place orders up to five years ahead. Why it matters: a hyperscaler moving inference onto its own chips is a direct read-across for merchant GPU demand (NVDA, AMD) and for HBM/foundry tightness (TSM, MU). What to watch: whether the GPU-comparison claims are verified, and HBM supplier disclosures.