Microsoft has announced Maia 300, its third in-house AI chip, and the pitch is that it will do most of the company's inference work so that Nvidia and AMD get fewer of its purchase orders. The catch is the calendar. The chip starts reaching production workloads in 2027, and its predecessor was only announced in January.
What Microsoft actually announced
Microsoft calls Maia 300 the "primary AI inference system" for its products and services. The company says it cuts the cost and raises the efficiency of running AI in the cloud, and that it suits the newer reasoning models and agentic tasks better than earlier Maia parts.
Scott Guthrie, executive vice president of Microsoft's cloud group, put two numbers on it:
- Efficiency: Microsoft expects 40 percent more tokens per megawatt than leading GPU-based systems.
- Scale-up: In a full 96-accelerator configuration, rack-scale token throughput reaches up to 7x that of Maia 200.
Note the verbs. The efficiency figure is something Microsoft "expects," not something it has measured in public. The comparison is also to "leading GPU-based systems" in general, with no particular product named. No benchmark methodology came with the announcement.
The plan to need Nvidia less
The strategic claim is bigger than the spec sheet. Guthrie said Microsoft will scale its Azure Maia family to multi-gigawatt deployments starting in 2027, with Maia as the primary inference system for Microsoft products and services.
That is a statement about inference, the business of running models for users, rather than about every AI workload Microsoft has. The framing of the announcement is that Microsoft can soon lean mainly on its own silicon instead of chips from Nvidia and AMD. "Soon" here means a rollout that begins next year, which is soon in chip years and slightly less soon in the budget year Nvidia is currently enjoying.
Maia 200 is already at work
Maia 300 succeeds Maia 200, and the gap between them is short. Maia 200 is already running production AI workloads across Microsoft, including Microsoft Copilot, GitHub Copilot, Microsoft Foundry and the company's own AI models.
Guthrie called Maia 200 the best performing AI infrastructure in Microsoft's fleet, measured in real-world metrics for AI services. That claim is useful context for the new chip. Microsoft is not asking anyone to take Maia on faith, because it says the earlier generation is already carrying real traffic.
Maia 300, in his description, is a complete AI system built for the workloads now emerging:
- long-context reasoning
- latency-sensitive conversational experiences
- sustained, multi-step agentic tasks
Chiplets and TSMC's N2P
The efficiency gains come largely from a new chiplet architecture. Maia 300 is built on TSMC's N2P process node, and it separates compute, memory controller and I/O functions into distinct pieces.
Microsoft's Azure hardware and infrastructure team says the split increases the silicon area available for compute and lets each function use the technology best suited to its role. Put plainly, the parts that do arithmetic no longer have to share a die with the parts that merely move data around, which is the sort of arrangement that sounds obvious once someone has paid to engineer it.
Chips as part of the model race
Maia is not only a cost-cutting exercise. Microsoft treats it as part of its own AI model work, and the announcement ties it to the goal of catching up with Google DeepMind, OpenAI and Anthropic.
Microsoft AI CEO Mustafa Suleyman said the company has to work across the full stack, down to the silicon. At the scale Microsoft operates, he said, it needs a platform it owns and co-designs with its models, which is why running Microsoft AI models on Maia is a critical focus for his team.
The logic is straightforward. If you design the chip and the model together, you can tune each around the other's quirks. If you rent someone else's chip, you tune around theirs.
What remains unproven
The announcement leaves several questions open. It does not say how large the first deployments will be, beyond the multi-gigawatt ambition from 2027. It does not say which GPU system the 40 percent comparison uses. It also says nothing about how much Nvidia and AMD hardware Microsoft will keep buying in the meantime.
What it does establish is a direction. Microsoft has a second-generation chip already in service, a third on TSMC's N2P node arriving in 2027, and an executive saying the in-house part will be the main inference engine. Whether the 40 percent holds up outside the press release will be answered by the production workloads, which is where chip claims eventually have to go and stay.