TLDR Tech / Launches
Microsoft's Maia 300 accelerator uses HBM4 and 167kW racks
Microsoft has introduced Azure Maia 300, its third-generation AI inference accelerator and first multi-chiplet design. It features an N2P compute die and eight HBM4 stacks totaling 288GB of memory. A rack of Maia 300 contains 96 accelerators within a 167kW power budget. Microsoft claims this system offers 40% more tokens per megawatt than leading GPU systems, with large-scale deployments planned for 2027.
Read the full story