Microsoft introduced its Azure Maia 300 chip, aiming to lessen dependence on Nvidia and AMD as hyperscalers push deeper into custom silicon. Unveiled at the OCP Global Summit, the Maia 300 is slated for multigigawatt deployments next year and is set to become Microsoft’s primary inference engine across products and services. The company says the chip delivers 40% more tokens per megawatt than leading GPU-based systems and enables a 96-chip rack to achieve seven times the token throughput of its Maia 200. The move underscores a broader industry shift as Amazon, Google and Meta build or rent out their own accelerators, pressuring Nvidia, which derives a large share of sales from cloud providers. Even so, analysts expect hyperscalers to keep buying Nvidia hardware to cover a range of AI workloads.
Related articles:
— Google’s Tensor Processing Unit (TPU) family overview
— Meta’s MTIA custom AI accelerator
— AMD Instinct accelerators, including MI300 series
— Open Compute Project and open hardware for hyperscale data centers
