Microsoft Unveils Next-Gen AI Chip to Boost Cloud Services

• The Maia 200 uses TSMC's 3nm process with four chips per server connected via Ethernet (not Nvidia's InfiniBand), and can scale up to 6,144 chips working together
• Microsoft's AI division led by Mustafa Suleyman, Microsoft 365 Copilot, and Microsoft Foundry service will use the chips, with wider customer availability planned for the future
• The chip packs more high-bandwidth memory than Amazon's third-gen Trainium or Google's seventh-gen TPU, targeting efficiency amid power consumption concerns in data centers

Microsoft Unveils Maia 200 AI Chip to Challenge Nvidia and Cloud Rivals

Microsoft announced its second-generation AI chip, the Maia 200, positioning it as a competitive alternative to Nvidia processors and cloud offerings from Amazon Web Services and Google. The company claims the chip delivers 30% higher performance than alternatives at equivalent pricing.

Key Technical Specifications:

• Built using TSMC's 3-nanometer process
• Four chips connected per server
• Uses Ethernet cables instead of Nvidia's InfiniBand standard
• Contains more high-bandwidth memory than AWS's third-generation Trainium or Google's seventh-generation TPU
• Supports clustering up to 6,144 chips for enhanced performance

Deployment and Availability:

The Maia 200 is currently rolling out in U.S. data centers, starting with the U.S. Central region, followed by U.S. West 3, with broader customer availability planned for the future. Microsoft's internal AI team, led by Mustafa Suleyman, will use the chips, along with Microsoft 365 Copilot for commercial clients and the Microsoft Foundry service.

Market Context:

This launch comes two years after Microsoft's first-generation Maia chip, which was never made available to cloud customers. The move addresses surging demand from generative AI developers like Anthropic and OpenAI, while helping data center operators balance computing power with energy efficiency.

Competitive Positioning:

By developing proprietary chips, Microsoft aims to reduce dependence on Nvidia while competing more effectively against AWS and Google, both of which offer custom AI processors. The chip's Ethernet-based architecture avoids reliance on Nvidia's InfiniBand technology, acquired through its 2020 Mellanox acquisition for $7 billion.