Microsoft details Maia 300 inference accelerator for Azure scale-up

Microsoft disclosed architectural details of its Maia 300 accelerator, a third-generation in-house AI chip engineered for inference workloads at Azure scale. The processor adopts a chiplet design and integrates HBM4 memory with high-bandwidth interconnects to raise inference efficiency across the cloud fleet.

The move gives Microsoft tighter control over its silicon roadmap, which can lower dependence on external suppliers and reduce operating costs over time. Broad deployment is not immediate; the company plans to roll out the Maia 300 in 2027.

Source: https://azure.microsoft.com