Tuesday, October 13, 2026

Microsoft unveils its own AI chip, 'Maia 300' ... 'At least 1 million to be deployed next year'

Input
2026-10-13 07:50:59
Updated
2026-10-13 07:50:59
Company sign and logo at Microsoft headquarters in Redmond, Washington State, US. AP Newsis

[Financial News] Microsoft (MS) unveiled its third-generation in-house artificial intelligence (AI) chip, 'Maia 300,' and said it would deploy it at scale starting next year to reduce its reliance on NVIDIA.
MS said in an official blog post on the 12th (local time) that, starting next year, it would scale up deployment of Azure Maia chips to several gigawatts (GW) and make them the primary AI inference system for its products and services.
MS said it could deploy at least 1 million Maia 300 chips next year.
Scott Guthrie, MS executive vice president overseeing its Cloud and AI division, said, "With the arrival of the superintelligence (SI) era, MS is entering a new phase of infrastructure," adding, "Maia 300 will process 40% more tokens per megawatt than leading graphics processing unit (GPU)-based systems."
MS did not disclose which systems it was using for comparison.
Unlike NVIDIA GPUs, in which memory surrounds a single massive compute chip, Maia 300 uses a 'multi-chiplet' design that bonds multiple specialized chips together.
The core compute chiplet uses Taiwan's TSMC 2-nanometer process, while chiplets responsible for data transfer and other tasks use a 3-nanometer process.
Each chip contains eight sixth-generation high-bandwidth memory (HBM4) units, delivering memory bandwidth of 18 terabytes (TB) per second, and has a thermal design power (TDP) of 950 watts.
A rack housing 96 accelerators draws about 167 kilowatts (kW), and its throughput is up to seven times that of its predecessor, Maia 200. NVIDIA's latest GB300 rack uses 142 kW to power 72 GPUs.
Maia 200, unveiled last January, is currently used in Microsoft Copilot and other products, and its deployment will expand to GitHub Copilot and others by year-end.
Cloud rivals such as Google and Amazon, as well as AI companies such as OpenAI, are also stepping up development of their own chips to reduce their reliance on NVIDIA. MS expects its capital expenditure this year to exceed $175 billion (about 237.7 trillion won).
MS plans to reveal more of its strategy at its annual developer event, 'Ignite 2026,' to be held in San Francisco next month. 


[email protected] Lee Seok-woo, international affairs specialist Reporter