Microsoft Unveils Maia 300, Its Third-Generation AI Chip, Aimed at Becoming Its "Primary Inference System" From 2027

Microsoft announced Maia 300, its third-generation in-house AI chip, on October 12, 2026. According to The Verge, Microsoft positions it as the "primary AI inference system" behind its products and services. Inference is the computing a model uses each time it answers a request after training is done. As chat, coding assistants and agents see heavier use, that spend has become one of the largest costs for cloud providers. Microsoft's move is to shift more of that work onto chips it designs itself.

The timeline comes first. Maia 300 will start rolling out to production in 2027, so it cannot be used today. It is a roadmap item. Its predecessor, Maia 200, was announced only in January and already runs production workloads across Microsoft, including Microsoft Copilot, GitHub Copilot, Microsoft Foundry and Microsoft's own AI models. Scott Guthrie, who leads Microsoft's cloud group, calls Maia 200 the best-performing AI infrastructure in the fleet when measured in real-world metrics for AI services. A gap of less than a year between generations suggests a fast cadence.

Now performance. Guthrie says Microsoft expects Maia 300 to deliver 40 percent more tokens per megawatt than leading GPU-based systems. In a full 96-accelerator configuration, he says it delivers up to 7x the rack-scale token throughput of Maia 200. Two points deserve care. Both figures are executive expectations, and the article cites no independent testing. And the unit matters: data centers are increasingly limited by electrical power rather than by the number of machines, so Microsoft frames the story around power efficiency instead of peak compute alone.

One architectural detail is concrete. Maia 300 is built on TSMC's N2P process node and uses a chiplet design that separates compute, memory controller and I/O functions. Microsoft's Azure hardware and infrastructure team says this increases the silicon area dedicated to compute and lets each function use the technology best suited to its role. Guthrie describes Maia 300 as a complete system for the new types of workloads: long-context reasoning, latency-sensitive conversational experiences and sustained, multi-step agentic tasks. That matches how products have shifted over the past year. Agents often call tools repeatedly across a long context, and they ask more of latency and stability than a single-turn answer does.

The strategic intent sits in one sentence. Microsoft will scale the Azure Maia family to multi-gigawatt deployments starting in 2027, making Maia the primary inference system for Microsoft products and services. From this, The Verge concludes that Microsoft can soon rely mainly on in-house chips rather than Nvidia's and AMD's. That is a reading of a public statement, not an announcement that Microsoft will stop buying outside silicon. Mustafa Suleyman, CEO of Microsoft AI, gave a second motive: the company has to work across the full stack, all the way down to the silicon, and at that scale it needs a platform it owns and co-designs with its models. Running Microsoft's AI models on Maia is a critical focus for his team.

Our view is that the story is about cost structure. If inference runs mostly on its own chips, Microsoft keeps more of the margin and can tune models and hardware together. The risk is just as clear. The chip arrives in 2027, and there is no evidence yet that delivery will be on time or that real-world efficiency will meet the expectations. The report also gives no memory capacity, peak compute, price, or word on whether outside cloud customers can use it. Those gaps cannot be judged until Maia 300 ships. Read it as a signal: large cloud providers are moving toward vertical integration of models and chips, and inference efficiency is becoming the battleground.