Microsoft unveils in-house Maia 100 AI chip and Cobalt 100 CPU

Microsoft announced its first custom-designed datacentre silicon at its Ignite conference: the Azure Maia 100, an AI accelerator built for training and running large language models, and the Azure Cobalt 100, an Arm-based general-purpose cloud CPU. Both chips were designed in-house rather than sourced from Nvidia or Intel, Microsoft’s existing suppliers for accelerators and server processors respectively.

Microsoft framed the move as a “systems approach,” arguing that designing silicon, servers, cooling and software together — rather than assembling a stack from off-the-shelf components — would let it tailor performance and efficiency specifically to its own AI and cloud workloads. The company said both chips would begin rolling out to its datacentres in early 2024, initially powering internal services including Microsoft Copilot and the Azure OpenAI Service, before any prospect of third-party customer access.

The announcement placed Microsoft alongside Google, Amazon and Meta, all of which had by then developed or were developing their own AI accelerators to reduce dependence on Nvidia’s GPUs — then the industry’s dominant and severely supply-constrained training hardware. For a company that was simultaneously OpenAI’s primary infrastructure partner and one of Nvidia’s largest customers, in-house silicon represented a hedge rather than a replacement: Microsoft continued to buy Nvidia and AMD chips at scale, and Maia’s near-term role was to absorb a slice of internal demand rather than to compete for the external accelerator market. The move set expectations for a widening pattern in which every major cloud provider would field its own custom chip alongside continued heavy reliance on Nvidia.