Microsoft Prepares to Mass-Produce Custom AI Chips for Azure

Microsoft is set to reveal its Maia 300 AI accelerator as early as September this year, expanding its custom silicon strategy. The Maia 300 will support both AI training and inference, with a key focus on scaling production beyond one million chips.

The company is partnering with TSMC, Taiwan Semiconductor Manufacturing Company, who delivered the first iteration of Maia chips. However, this time the scale is much larger, increasing from tens of thousands of Maia 200 accelerator units to potentially hundreds of thousands or even over a million Maia 300 chips.

Microsoft's goal with Maia 300 is not to compete directly with NVIDIA but rather ensure that Azure doesn't rely entirely on NVIDIA infrastructure. By investing in these custom chips, Microsoft aims to have greater control over the cost, supply, and performance of AI infrastructure running on Azure.