Microsoft has introduced its newest chip, the Maia 200, which it calls a robust silicon solution built to handle the scaling of AI inference. This release follows the Maia 100 from 2023 and is engineered to operate advanced AI models more quickly and efficiently.

The Maia 200 incorporates over 100 billion transistors, achieving over 10 petaflops in 4-bit precision and around 5 petaflops in 8-bit performance, marking a significant upgrade from the earlier model. Inference, the computational task of executing a trained model, has grown into a major operational expense for AI firms as they evolve, driving a focus on optimizing this process.

Microsoft aims for the Maia 200 to contribute to that optimization, enabling AI operations to run more smoothly with reduced power consumption. The company stated, “In practical terms, one Maia 200 node can effortlessly run today’s largest models, with plenty of headroom for even bigger models in the future.”

This launch aligns with a broader movement among major tech companies to develop custom chips, reducing reliance on NVIDIA’s GPUs, which have become essential for AI advancements. For example, Google offers its Tensor Processing Units (TPUs) as cloud-based compute power, while Amazon recently debuted the Trainium3, the latest iteration of its AI accelerator chip, in December.

These alternatives help shift some computational workloads away from NVIDIA hardware, cutting overall costs. Microsoft is positioning Maia to compete directly, claiming it delivers three times the FP4 performance of Amazon’s third-generation Trainium chips and surpasses Google’s seventh-generation TPU in FP8 performance.

Already, Maia is actively powering AI models from Microsoft’s Superintelligence team and supporting its Copilot chatbot. As of this week, the company has opened access to its Maia 200 software development kit for various users, including developers, academics, and leading AI labs, to integrate into their own workloads.