Microsoft's Maia 300 rack: 96 accelerators, HBM4, mostly liquid-cooled

Microsoft's real bet isn't the chip — it's the rack. 96 accelerators per cabinet, two-level networking to 12,288 units: the unit of AI competition has shifted from one GPU to one rack. This is inference economics forcing a structural redesign, not a spec bump. HBM4 is the enabler that makes 96-card density physically viable under thermal and power ceilings. Chain reaction: HBM4 drags SK Hynix and Samsung into a new yield war. Liquid cooling graduates from differentiator to entry requirement — Vertiv, Envicool, and peers see TAM expand 3–5x. The two-level interconnect implies custom network silicon, eroding NVLink's gravitational pull. Risk: The four-layer closed stack (software + chip + rack + DC) is simultaneously a moat and a single point of failure. Under HBM4 export controls, 12K-cluster delivery becomes hostage to CoWoS yield and capacity allocation in Taiwan, China. Competition: NVIDIA's GB300 NVL72 is the head-on rival, but Microsoft is optimizing inference TCO, not peak FLOPS. AMD MI400 and Google TPU v6 must match rack-level density or lose the agentic-inference segment. NVIDIA's existential threat isn't a faster chip — it's a better rack. 12–24 months: liquid cooling becomes default above 500 kW; rack-level co-design replaces chip-level competition as the dominant paradigm; inference bifurcates — training stays with NVIDIA, inference migrates to custom silicon.