Core Points

• Microsoft stated that in the future, the deployment scale of its Maia series AI chips will reach several gigawatts of computing power.
• The company plans to use Maia as the main chip architecture for running the AI model.
• The next generation of Maia 300 chips will be put into production in 2027.

After years of relying on NVIDIA graphics processors, Microsoft will expand the use of its own developed artificial intelligence chips. This software giant aims to strengthen its own full-stack AI capabilities.

Microsoft officially released its third-generation Maia AI acceleration chip on Monday, stating that it will rely primarily on its own chip hardware for handling business requests in the future.

Microsoft's greatest advantage in the AI field is its Azure cloud platform, which runs various AI models for customers. Microsoft 365 Copilot and GitHub Copilot provide AI services for programmers and office workers. Microsoft is also developing its own large models, but in a market dominated by OpenAI, Anthropic, and Google, its self-developed models have not yet gained significant influence.

The chip industry is facing new opportunities, driven by the persistent strong demand for high-end processors and the market's desire for a variety of chip options. NVIDIA GPU has always been the industry standard for computing power and model training in AI. Microsoft's existing data centers make extensive use of this product, which is also used to run models for OpenAI and Anthropic.

However, as demand continues to rise, leading model companies have begun to use chips from multiple manufacturers, including Amazon, AMD, Cerebras, Google, and in recent months, Microsoft's self-developed chips have also been incorporated.

In general computing power scenarios, Microsoft has been using its own CPU as well as Intel's and AMD chips simultaneously. Chief Financial Officer Amy Hood stated at a Goldman Sachs conference last month that the company's in-house developed chips have contributed to improving the company's profit margins over the past decade.

Microsoft's first AI chip Maia 100 was launched in 2023; followed by the release of Maia 200. CEO Satya Nadella stated to analysts during a financial report conference call that Maia 200 "offers 30% more performance per dollar invested" than existing hardware, and this chip has already been used in OpenAI models.

Maia 300 is set to enter mass production in 2027. Scott Guthrie, Executive Vice President of Microsoft, stated in a blog post that the goal is to achieve a deployment scale of several gigawatts for the Maia series of chips, but General Manager Alistair Spears did not provide a clear timeline for implementation.

Spillers said, "We reserve a certain degree of flexibility in our choices. This industry is evolving at an extremely fast pace, and we hope to match the most suitable chips with the corresponding models at the right time."

Microsoft plans to deploy 96 Maia 300-chip units in each data center cabinet, which is twice the number of Maia 200 chips. The token processing throughput per cabinet can reach up to 7 times that of the previous generation; compared to GPU, the number of tokens produced per megawatt of energy consumption has increased by 40%.

"We need a brand-new architecture to achieve better performance and at the same time improve power utilization efficiency. For data centers, power consumption is directly correlated with operating costs," said Spears.

Microsoft Technical Fellow Jim Clivin stated that some of Microsoft 365 Copilot services can be accelerated with the help of Maia 300.

Analysts from investment bank Evercore ISI estimated in a research report sent to clients on Wednesday that by the first quarter of 2027, Maia chips will account for 10% of Microsoft's AI chip procurement expenses; institutions have given Microsoft's stock a buy rating.

NVIDIA will still be an important partner for Microsoft. Last week, during Microsoft's product launch event, NVIDIA CEO Jensen Huang was present. At the event, Microsoft unveiled new Surface laptops equipped with NVIDIA processors. Spears also mentioned that in the field of model training, Microsoft will continue to collaborate with NVIDIA and AMD.

Unlike Google’s TPU, Amazon’s Inferentia and Trainium chips, Microsoft’s acceleration chip is not currently available for rental on an hourly basis to other companies.

Patrick Mulheard, CEO of industry consulting firms Moor Insights and Strategy, stated: "As of now, apart from within Microsoft, no external companies have conducted any testing on this chip."