Microsoft unveils Azure Maia 300 for massive 2027 AI expansion

Microsoft is making a much bigger bet on its own AI hardware. The company announced plans on October 12 to expand its Azure Maia systems to multi-gigawatt deployments starting in 2027, with Maia becoming the primary inference system for Microsoft products and services.

The next step is Azure Maia 300, a new generation of Microsoft’s custom AI infrastructure scheduled to begin production deployments in 2027. Microsoft expects the system to deliver 40% more tokens per megawatt than leading GPU-based systems, while a full 96-accelerator configuration promises up to seven times the rack-scale token throughput of Maia 200.

Those are Microsoft’s projections, not independently verified benchmark results. Still, the direction is clear: Microsoft wants more of the AI behind its services running on hardware designed specifically for those workloads.

Azure Maia 200 is already powering Microsoft Copilot

This isn’t just a roadmap announcement. Microsoft says Maia 200 is already serving production workloads, including Microsoft Copilot, and supporting the Microsoft AI team as it develops next-generation models.

GitHub Copilot and Microsoft Foundry are expected to follow by the end of 2026.

Maia 200 supports OpenAI models, Microsoft’s own MAI models, and leading open-source model families. That makes the hardware part of a broader, model-diverse strategy rather than a platform built around one particular AI model.

For readers trying to keep Microsoft’s AI products straight, our Microsoft Copilot vs. Microsoft 365 Copilot guide explains the differences between the consumer assistant and the work-focused service. Maia is the infrastructure story underneath those kinds of experiences—not another Copilot subscription or app.

Microsoft says Maia 200 is deploying across the United States, Europe, and Australia. For built-in AI capabilities in Teams, Copilot, and Cowork experiences, the company reports up to 1.7 times more tokens per watt than leading GPU infrastructure.

The important qualifier is “up to.” That figure applies to the workloads Microsoft describes, not every AI model or task.

What makes Azure Maia 300 different?

Microsoft designed Maia 300 around AI workloads that demand more than a quick answer to a short prompt.

The company specifically points to long-context reasoning, responsive conversational experiences, and sustained, multistep agentic tasks. In plain English, that means AI that can work through more information, respond without frustrating delays, and keep carrying out a task across multiple steps.

That focus fits the shift we’ve covered in Microsoft’s Copilot leadership and agentic AI push, where the company is working toward a more connected Copilot system spanning experiences, platforms, productivity apps, and models.

Rather than focusing only on raw compute, Maia 300’s design emphasizes how efficiently information moves between compute, memory, and networking resources.

Microsoft says its software-defined dataflow architecture and AI-assisted kernel optimization reduce unnecessary data movement and improve utilization. The goal is to get more useful AI work out of the available hardware and electricity.

Azure Maia 300 specifications at a glance

Here are the headline specifications Microsoft provided in its announcement:

Component	Microsoft’s announced specification
Production deployment	Begins in 2027
Architecture	Inference-first, multi-chiplet system
Compute manufacturing process	TSMC N2P
Memory	Eight HBM4 stacks per accelerator
Peak memory bandwidth	18.0 TB/s per accelerator
Accelerator thermal design power	950 watts
Full rack configuration	96 accelerators
Rack power envelope	167 kilowatts
Rack HBM capacity	27.6 TB
Aggregate rack HBM bandwidth	1.7 PB/s
Expected efficiency versus leading GPU-based systems	40% more tokens per megawatt
Rack-scale throughput versus Maia 200	Up to 7 times higher

The accelerator and rack power figures describe different levels of the system. The 950-watt figure is the accelerator’s thermal design power, while 167 kilowatts is the power envelope Microsoft gives for the full rack system.

Microsoft also says Maia 300 supports advanced low-precision formats, including enhanced MXFP4, to reduce memory and bandwidth requirements for computation and KV-cache storage.

Its networking design uses Ethernet-based infrastructure with a custom AI transport to support communication across accelerators and racks. Microsoft intends that architecture to scale from a single rack to datacenter-scale deployments.

Why Microsoft is focusing on tokens per megawatt

The standout number isn’t just the sevenfold rack-throughput claim. It’s Microsoft’s expectation of 40% more tokens per megawatt compared with leading GPU-based systems.

Microsoft is presenting Maia 300 as a way to deliver more AI output within a given power budget. That matters to its expansion plans because the announcement is about infrastructure measured in gigawatts, not a small hardware refresh.

The two performance claims also measure different things:

• The 40% figure compares expected power efficiency with leading GPU-based systems.
• The up-to-sevenfold figure compares rack-scale token throughput with Maia 200.

Neither figure means every Copilot response will become seven times faster. The announcement also does not name the GPU systems used for comparison or provide enough benchmark detail to make a universal performance judgment.

Which Microsoft services will use Maia?

Microsoft describes Maia 200 and Maia 300 as supporting a broad range of AI services, including:

• Microsoft Copilot and built-in AI experiences across Teams and Cowork.
• GitHub Copilot, including more efficient multistep coding experiences.
• Microsoft Foundry, supporting different model architectures and deployment environments.
• Microsoft AI’s model inference and data-generation work.
• Dragon Copilot, supporting latency-sensitive clinical workflows.

For Microsoft 365 users, this is the infrastructure side of the broader effort to make AI more useful inside everyday apps. Our Microsoft 365 Copilot overview covers how those experiences fit into Word, Excel, PowerPoint, Outlook, and Teams.

The announcement does not promise an immediate subscription price cut, a specific new Copilot feature, or a guaranteed response-time improvement. It sets out how Microsoft plans to supply the computing capacity behind its growing AI services.

Maia 300 deployments begin in 2027

Microsoft says Azure Maia 300 will begin production deployments in 2027, alongside the start of its broader multi-gigawatt Maia expansion.

Making Maia the primary inference system for Microsoft products and services would give the company a larger role in shaping the hardware behind its AI experiences. But “primary” does not mean “exclusive”: the announcement also emphasizes hardware and model choice across a heterogeneous platform.

For now, Maia 200 is already doing production work. Maia 300 is the next planned step, with a clear goal: handle larger, longer-running AI workloads while getting more useful output from every megawatt. Check out the Maia 300 site.