---
title: "Microsoft details Maia 300, its third-generation in-house AI inference chip"
url: https://anawords.com/story/microsoft-maia-300-inference-chip
published: 2026-10-13T10:30:00+05:30
updated: 2026-10-13T10:30:00+05:30
section: Chips & Compute
author: Mayank Sahu
checked: sources checked by AI, approved by editor
version: 1
---

# Microsoft details Maia 300, its third-generation in-house AI inference chip

Microsoft says the chip, built on TSMC’s N2P process with 288 GB of HBM4, gives up to seven times the rack-level token throughput of Maia 200. The performance figures are Microsoft’s own.

**In short:** Microsoft published details of Maia 300, its next in-house AI inference accelerator, on 12 October 2026. Microsoft says it uses TSMC N2P compute dies and 288 GB of HBM4, and claims up to 7x the rack-scale token throughput of Maia 200.

## What Microsoft announced

Microsoft published a technical description of Maia 300, its next in-house AI accelerator, on 12 October 2026. The post is written by Rani Borkar, president of Azure Hardware Systems and Infrastructure, and Saurabh Dighe.[^1] Microsoft says the chip is built for “agentic inference”, including long-context and reasoning-heavy workloads.[^1]

## The design

According to Microsoft, the compute dies are made on TSMC’s N2P process, with memory-controller and I/O functions on separate dies.[^1] Each accelerator has eight HBM4 stacks with 288 GB of capacity and 18 TB/s of bandwidth.[^1] SiliconANGLE notes that the capacity matches Nvidia’s Rubin accelerator and that the bandwidth is about 20% lower.[^2]

A rack holds 96 accelerators in 24 trays, and up to 128 racks can be joined into a superpod of 12,288 accelerators.[^1][^2] Microsoft says liquid cooling removes more than 80% of each rack’s heat.[^1]

## Microsoft’s performance claims

Microsoft claims up to 7x the rack-scale token throughput of Maia 200 and more than 2.5x the performance per watt.[^1] SiliconANGLE reports that Microsoft also claims 40% more tokens per megawatt than “leading GPU-based systems”.[^2] These figures have not been independently tested.

## Where it will be used

Microsoft says Maia 300 is designed to add capacity for Microsoft Copilot, GitHub Copilot and Microsoft Foundry.[^1] SiliconANGLE reports that Microsoft plans to begin production deployment in 2027 and expects the Maia series to become its main source of inference capacity over time.[^2] Microsoft’s post gives no availability date and says more will be shared at Microsoft Ignite, 17–20 November.[^1]

## Why it matters

Like Google and Amazon, Microsoft is building its own AI chips alongside buying from Nvidia and AMD. Maia 200 is already in production, according to SiliconANGLE.[^2]

## Fact ledger

- [Confirmed] Maia 300 compute dies use TSMC N2P; each chip has 288 GB of HBM4 at 18 TB/s. (source 1)
- [Confirmed] Racks hold 96 accelerators; superpods scale to 12,288 accelerators. (source 1,2)
- [Company claim] Up to 7x the rack-scale token throughput of Maia 200. (source 1)
- [Company claim] 40% more tokens per megawatt than leading GPU-based systems. (source 2)
- [Unconfirmed] Production deployment to begin in 2027. (source 2)
- [Not known] Pricing and Azure availability for customers.

## Sources

1. [Microsoft Azure Blog: Deep dive into Azure Maia 300 architecture](https://azure.microsoft.com/en-us/blog/deep-dive-into-azure-maia-300-architecture/) — official
2. [SiliconANGLE: Microsoft debuts multichiplet Maia 300 inference accelerator](https://siliconangle.com/2026/10/12/microsoft-debuts-multichiplet-maia-300-inference-accelerator/)