Microsoft's Maia 300 chip aims to beat GPUs on efficiency
Microsoft just unveiled Azure Maia 300, its next-generation AI inference chip set for multi-gigawatt rollouts in 2027. Featuring eight HBM4 stacks and 18 TB/s of bandwidth, it claims 40% more tokens per megawatt than leading GPU systems. Why it matters: Hyperscalers are betting custom silicon will overcome memory bottlenecks and rein in the staggering cost of serving long-context AI agents.