Microsoft Unveils Maia 300 AI Chip to Cut Its Reliance on Nvidia

Microsoft has unveiled Maia 300, the third generation of its in-house AI chip, and says the Maia family will become the main engine behind its AI products starting in 2027. The company presented the chip on October 12 at the OCP Global Summit in San Jose.

The headline claim: Maia 300 processes 40% more tokens per megawatt than "leading GPU-based systems." If that holds up in production, it changes the economics of running Copilot and other AI services at Microsoft's scale.

Why Microsoft is building its own chips

Every major cloud company is trying to make its own AI silicon. Amazon, Alphabet, and Meta all have custom chips, and OpenAI and Anthropic are working on their own as well.

The reason is simple: cost and control. Buying Nvidia and AMD accelerators at enormous scale is expensive, and supply is never guaranteed. Microsoft expects to spend roughly $175 billion on capital projects in calendar 2026, according to its last earnings call, much of it on AI hardware.

Maia 300 is Microsoft's clearest attempt yet to claw some of that spend back. Reuters reported that the rollout could involve more than one million Maia 300 chips next year.

What Maia 300 actually is

Maia 300 is an inference chip (built to run trained AI models and answer requests, rather than to train new models from scratch). Microsoft says it targets long-context reasoning, conversational AI, and multistep agents.

The key specs:

• Memory: eight stacks of HBM4 (high-bandwidth memory stacked right next to the processor), totaling 288 GB per chip
• Bandwidth: 18 TB/s per chip, which is 2.5x Maia 200
• Manufacturing: compute chiplets built on TSMC's N2P process
• Power: a 950 W budget per chip

Why this matters: modern AI models are often starved for memory, not raw math. Big, fast memory lets one chip hold longer conversations and serve more users at once.

The rack is the real product

Microsoft isn't just shipping a chip. It's shipping a full rack design built around it.

Each Maia 300 rack holds 96 accelerators in a 167 kW power envelope, with 27.6 TB of high-bandwidth memory. That's twice the accelerator density of Maia 200 racks. Microsoft says one rack can deliver up to seven times the token throughput of a Maia 200 rack.

For comparison, Reuters noted that Nvidia's GB300 racks pack 72 GPUs at about 142 kW. Liquid cooling handles more than 80% of the Maia 300 rack's heat, and the racks can run in both water-cooled and air-cooled data centers.

Built on a chip that's already working

This isn't a science project. Maia 200 already runs production workloads, including Microsoft Copilot, OpenAI models, and Microsoft's own MAI models.

GitHub Copilot and Microsoft Foundry are slated to move onto Maia 200 by the end of 2026. Microsoft says Maia 200 already delivers up to 1.7x more tokens per watt than leading GPU infrastructure for some Teams and Copilot features.

Who this affects

Nvidia and AMD face the most obvious pressure. Microsoft is one of their biggest customers, and every Copilot request that runs on Maia is one that doesn't run on their hardware. Nvidia and AMD shares fell on Monday while Microsoft's rose.

Azure customers could benefit if lower internal costs show up as cheaper inference. That's the real prize: more intelligence per dollar.

The wider industry gets more confirmation that inference, not training, is where the next hardware battle is being fought.

What insiders are saying

Rani Borkar, president of Azure hardware systems and infrastructure, framed the goal to Reuters as "delivering more intelligence for every watt, every dollar."

Analysts were more measured. They note that cloud providers are unlikely to stop buying Nvidia chips entirely, because different workloads need different processors. Microsoft itself announced plans in July to expand Azure with AMD's Helios platform, and Nvidia GPUs remain central to its fleet.

What's next

Production deployments begin in 2027, scaling to what Microsoft calls multi-gigawatt levels. In August, The Information reported Microsoft was negotiating with TSMC for more than 300,000 Maia 300 chips for 2027.

Microsoft plans to share more at its Ignite conference in San Francisco, November 17 to 20. Watch for independent benchmarks, because Microsoft hasn't named the GPU systems behind its 40% claim, and it hasn't published cost-per-token comparisons.

The bottom line

Maia 300 is Microsoft's bet that owning the chip, the rack, and the software beats renting someone else's. You won't buy one, but if you use Copilot, Azure AI, or OpenAI models through Microsoft, there's a growing chance your requests will run on it. The 40% efficiency number is the claim to watch, and 2027 is when we'll see if it's real.