Microsoft Maia 300 AI Chip: Specs, Rack Details, and Release Plan

Quick Answer

• Microsoft announced Maia 300 on October 12, 2026. It is the third generation of its in-house AI chip.
• It uses eight HBM4 memory stacks with 18 TB/s bandwidth, draws 950 W, and uses TSMC N2P for its compute chiplets.
• A rack of 96 chips has 27.6 TB of HBM and, Microsoft says, up to 7x the token throughput of a Maia 200 rack.
• Production deployments start in 2027, then scale to multi-gigawatt levels.
• Maia 300 is built for inference, which means running AI models, not training them.

What is Maia 300?

Maia 300 is Microsoft’s newest custom AI accelerator for Azure. Microsoft designs the chip, and TSMC makes the compute chiplets. Microsoft says Maia is expected to become the main inference hardware for its own products and services.

Inference is the step where a trained AI model answers your question. Chips built for it focus on memory speed and energy use, which is what Maia 300 stresses.

Maia 300 specs

Microsoft did not state the HBM capacity per chip in the sources we checked. Techzine’s rack figures imply about 288 GB per chip, but that is a calculation, not an official number.

Maia 300 rack details

Microsoft says a full Maia 300 rack reaches up to 7x the rack-scale token throughput of a Maia 200 rack. It also claims about 40% more tokens per megawatt than leading GPU-based systems. Both are Microsoft’s own claims, and independent tests are not available yet.

Maia 300 vs Maia 200

Bandwidth is more than double, and power is 200 W higher.

When will Maia 300 be available?

The first production deployments are planned for 2027. Microsoft says it will then scale Maia to multi-gigawatt levels. The reports do not say that outside customers can rent Maia 300 yet, so treat it as an internal Microsoft platform for now.

What will Maia 300 run?

According to IT Brief, Microsoft is aiming Maia 300 at:

• Multi-step agent systems, such as GitHub’s coding tools.
• Cloud-scale model serving in Microsoft Foundry.
• Inference and data generation for Microsoft’s MAI models.
• Latency-sensitive experiences, such as Dragon Copilot.

Maia 200, the current chip, already serves Microsoft Copilot and runs OpenAI models, Microsoft’s MAI models, and open-source models. GitHub Copilot and Foundry are due to move to Maia 200 by the end of 2026. For a simple explanation of agents, read our guide on AI agents.

Why does Microsoft build its own chips?

Companies that run AI at huge scale spend a great deal on chips and power. A custom chip can lower the cost per answer and reduce dependence on outside suppliers. It also fits the push for more efficient data centers, which is a growing topic. See our reports on Micron and Intel in the AI chip race, Google’s nuclear power deal, and Microsoft and Meta cutting Claude spending.

What does it mean for you?

Nothing changes today. If you use Copilot, the results may get faster or cheaper over time as Microsoft moves more work to its own chips. Developers who use Azure should watch for announcements on access and pricing.

FAQs