Microsoft unveils game-changing AI inference chip set to revolutionize machine learning performance

DCM Editorial Summary: This story has been independently rewritten and summarised for DCM readers to highlight key developments relevant to the region. Original reporting by Tech Crunch, click this post to read the original article.

Microsoft has introduced its newest in-house AI chip, the Maia 200, designed to improve the performance and efficiency of running powerful AI models. Following the Maia 100 released in 2023, the new chip boasts more than 100 billion transistors and delivers over 10 petaflops of 4-bit precision computing. This advancement makes the chip capable of handling even the largest AI models with ease, offering enough capacity to support even more complex models in the future.

If you’re working in AI, you know that inference—the process of running trained models—is becoming a significant operational cost. Microsoft aims to reduce that burden for AI businesses with the Maia 200, which promises not only increased speed and efficiency but also reduced power consumption. The chip is part of Microsoft’s broader strategy to optimize AI workflows and reduce disruptions in AI deployments.

By developing Maia in-house, Microsoft is joining other tech giants like Google and Amazon in an effort to move away from reliance on NVIDIA’s GPUs. Google uses its own TPUs (tensor processing units) and Amazon offers Trainium chips, with the latest Trainium3 launched recently. Microsoft claims that Maia outperforms both, offering three times the FP4 performance of Amazon’s Trainium3 and greater FP8 computing power than Google’s seventh-generation TPUs.

Currently, the Maia 200 is helping to power Microsoft’s Superintelligence team efforts and the operation of its AI Copilot assistant. To support broader adoption, Microsoft has also released a software development kit for the Maia 200, which you’re now invited to explore—especially if you’re a developer, academic, or working in a frontier AI lab.