Microsoft unveils Maia 300 chip, targets AI inference dominance by 2027

Microsoft announced the Maia 300, its third-generation custom AI chip, on Monday. The company plans to transition its AI services to this new silicon by 2027.

What happened

Microsoft unveiled the Maia 300 during a Monday announcement, outlining a strategy to shift its AI infrastructure toward proprietary hardware. The company intends for these chips to become the primary system for running artificial intelligence in its products starting in 2027. This move follows the launch of the second-generation Maia 200, which is already active in Microsoft’s data centers. The new chip is specifically engineered for inference, the process of executing trained AI models rather than building them. Microsoft stated the hardware is optimized for long conversations, complex reasoning, and autonomous agents that perform multi-step tasks.

The engineering design splits the processor into smaller components known as chiplets. These handle computing, memory, and data transfer separately. The computing sections utilize TSMC’s N2P manufacturing process. Each Maia 300 unit includes eight stacks of HBM4 memory. This architecture allows for greater efficiency in power and memory usage. Microsoft executives described the design as optimizing intelligence across every watt, dollar, and byte of memory. Rani Borkar, who leads Azure hardware systems, emphasized this holistic approach to performance gains.

Deployment scales are expected to be massive. Microsoft plans to deploy infrastructure measured in gigawatts starting next year. This could involve more than a million Maia 300 chips by 2027. A single rack containing 96 chips draws approximately 167 kilowatts of power. In comparison, Reuters reported that Nvidia’s latest GB300 racks hold 72 GPUs and consume about 142 kilowatts. Microsoft claims each megawatt of power yields 40% more tokens with Maia 300 than with top GPU systems. However, the company did not specify which systems it used for this comparison.

Why it matters

This strategic shift reflects a broader industry trend among cloud giants to reduce dependency on Nvidia. Google and Amazon already operate their own AI chips. OpenAI has also partnered with Broadcom to develop custom silicon. Microsoft’s move consolidates its position in this competitive landscape. By controlling the hardware stack, the company can optimize performance for its specific software needs. The Maia 200 already serves Microsoft Copilot and runs OpenAI models alongside Microsoft’s own MAI models. Support for GitHub Copilot and Microsoft Foundry is expected by the end of the year.

Securing manufacturing capacity is a critical part of this plan. In August, The Information reported that Microsoft was in talks with TSMC. The discussions involved producing more than 300,000 Maia 300 chips for the 2027 target. This volume underscores the scale of the upcoming infrastructure overhaul. Microsoft plans to share additional technical details at its Ignite conference in San Francisco next month. The transition signals a long-term commitment to custom silicon for AI inference.

What to watch

• Monitor deployment timelines for Maia 300 in Azure data centers.
• Track performance benchmarks against Nvidia’s latest GPU offerings.
• Observe capacity agreements between Microsoft and TSMC for 2027 production.
• Review updates on supported AI models and services using the new chips.

What to do and how to stay safe: Microsoft

• Assess current AI workload dependencies on third-party hardware providers.
• Evaluate long-term cost structures for inference services as custom chips scale.
• Monitor vendor announcements for changes in service level agreements or pricing.
• Plan infrastructure strategies that allow flexibility between custom and standard silicon.

Step-by-step guide: Software Updates Best Practices: Secure Patching Without Downtime

General security guidance from the Patch Gazette newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

When will Microsoft start using Maia 300 for its AI services?

Microsoft plans to make Maia 300 the primary system for running AI in its products starting in 2027.

How does Maia 300 compare to Nvidia’s hardware in terms of power and output?

A Maia 300 rack holds 96 chips and draws 167 kilowatts, while Microsoft claims it generates 40% more tokens per megawatt than top GPU systems.

What AI models currently run on Microsoft’s custom chips?

The existing Maia 200 chips serve Microsoft Copilot, run OpenAI models, and support Microsoft’s own MAI models.