Microsoft has unveiled the Maia 300, a third-generation AI accelerator featuring 288 GB of HBM4 memory and 18 TB/s of bandwidth. Built on TSMC's N2P process with a multi-chiplet design, the 950W chip delivers up to 13.8 PFLOPS of FP4 compute performance.
The hardware succeeds the 2026 Maia 200, offering 6.9 PFLOPS at FP8 precision. Utilizing eight HBM4 stacks and a two-tier Ethernet network, the architecture supports clusters of 12,000 units. Production deployment for the new silicon is scheduled for 2027.
This launch targets massive scale, with Reuters reporting potential deployment of one million chips. Microsoft claims the Maia 300 provides 40% more tokens per megawatt than top GPUs, while a 96-chip rack delivers seven times the throughput of its predecessor.