Microsoft used a 12 October post from cloud chief Scott Guthrie to introduce Azure Maia 300, its third-generation in-house AI accelerator, entering production in 2027. From that point Microsoft says the Maia family becomes the primary inference system for its own products and scales to multi-gigawatt deployments.
The claimed numbers: 40% more tokens per megawatt than leading GPU-based systems, 288GB of memory and 18TB/s of bandwidth — 2.5x its predecessor — and a 96-accelerator rack reaching up to 7x the rack-scale token throughput of Maia 200. Maia 200 is already in production in the US, Europe and Australia, running Teams, Copilot and Cowork at up to 1.7x tokens per watt, with GitHub Copilot and Microsoft Foundry joining by year end.
Alongside it came Cobalt 200, an agent-native CPU, claiming more than 50% higher performance than Cobalt 100, 33% lower agent call latency and 23% higher agentic-workload throughput. Microsoft AI CEO Mustafa Suleyman framed the whole thing as having to "work across the full stack, all the way down to the silicon." The subtext is the one everyone reads: less Nvidia and AMD in Microsoft's own inference path.
Sources:
Microsoft ·
Yahoo Finance