Microsoft's Maia 300 chip aims to become Azure's primary AI inference system from 2027

Microsoft targets 2027 for Maia 300 production

Microsoft has unveiled Maia 300, the third generation of its in-house AI accelerator, with an ambitious goal attached: becoming the main inference engine behind the company's cloud products and services. According to The Verge, the chip will start rolling out to production in 2027 and is intended to anchor multi-gigawatt-scale Azure deployments.

Scott Guthrie, executive vice president of Microsoft's cloud group, says the new accelerator delivers up to 7x the rack-scale token throughput of Maia 200 when built out in a full 96-accelerator configuration, and that Microsoft expects it to achieve "40 percent more tokens per megawatt compared to leading GPU-based systems." That framing matters: tokens per megawatt measures how much AI output a fixed amount of power buys, and power availability has become one of the binding constraints on data center growth.

There is also a supplier dimension. As The Verge reports, the efficiency gains are meant to let Microsoft lean primarily on its own silicon rather than GPUs from Nvidia and AMD. Guthrie says the company will "scale our Azure Maia family to multi-gigawatt deployments starting in 2027, making Maia the primary AI inference system for Microsoft products and services."

Chiplets on a cutting-edge process node

Maia 300 moves to a chiplet design built on TSMC's N2P manufacturing process, per The Verge. Instead of a single monolithic die, the architecture splits compute, memory controller, and I/O functions into separate pieces. Microsoft's Azure hardware and infrastructure team says this raises the share of silicon dedicated to compute and lets each function use the technology best suited to its role.

The chip is also aimed squarely at newer workload patterns. Guthrie describes it as a complete AI system built for long-context reasoning, latency-sensitive conversational experiences, and sustained, multi-step agentic tasks — the kinds of jobs that generate heavy, ongoing inference demand rather than one-off training runs.

Maia 200 is already carrying production traffic

The announcement continues a fast cadence. Maia 300's predecessor, Maia 200, was only announced in January, and it is already serving production workloads across Microsoft, including Microsoft Copilot, GitHub Copilot, Microsoft Foundry, and Microsoft's own AI models. Guthrie calls Maia 200 the best-performing AI infrastructure in Microsoft's fleet when judged by real-world metrics for AI services. That claim comes from Microsoft itself; the company has not offered independent benchmarks for the comparison.

Silicon as a lever for Microsoft's model ambitions

The Verge also connects the chip to Microsoft's frontier-model strategy. Mustafa Suleyman, CEO of Microsoft AI, argues the company has to "work across the full stack, all the way down to the silicon," and that at scale it needs a platform it owns and co-designs with its models. He calls running Microsoft AI's models on Maia a critical focus for his team. The report suggests custom silicon could help Microsoft close the gap with rivals such as Google DeepMind, OpenAI, and Anthropic.

Why it matters

Microsoft is pushing further down a path already trodden by Google and Amazon: hyperscalers designing their own inference silicon at scale. If Maia 300 meets its efficiency targets, Microsoft gains cost control over the most expensive part of running AI services and real negotiating leverage against Nvidia, whose GPUs currently define AI compute.

The emphasis on inference rather than training is also telling. As agentic, multi-step workloads multiply, serving models becomes the dominant volume of compute demand, and the company that owns an efficient inference platform shapes both its margins and its pricing. For Azure customers, better tokens-per-megawatt economics could eventually translate into cheaper or faster AI services.

The caveats are real, though. The headline numbers are Microsoft's own, production does not start until 2027, and Maia 300 will have to prove itself against incumbent GPUs outside vendor talking points before the Nvidia dependency actually loosens.