Microsoft moves its AI services to proprietary chips to reduce dependence on Nvidia
American corporation Microsoft announced a large-scale deployment of its own processors for artificial intelligence Azure Maia. Starting in 2027, the company plans to make them the main computing platform for AI operations in its products and services, reducing dependence on expensive Nvidia accelerators.
Executive Vice President of Microsoft for cloud technologies and artificial intelligence Scott Guthrie reported this in the official company blog on October 12.
Microsoft plans to deploy Azure Maia infrastructure with a capacity of several gigawatts and begin using the new generation of Maia 300 processors in 2027. This primarily concerns computations necessary for running already trained artificial intelligence models, particularly generating responses to users.
The company expects that its own hardware will reduce costs for processing AI requests, increase data center energy efficiency, and gain more control over technological infrastructure.
New chips should be more efficient than graphics processors
The key element of the strategy will be Azure Maia 300 - the third generation of Microsoft's specialized AI accelerators.
According to company estimates, Maia 300 will be able to generate 40% more tokens per megawatt of consumed electricity compared to leading systems based on graphics processors.
Tokens are units of text that language models process and generate. A higher number of tokens per unit of consumed energy means the ability to serve more requests without a proportional increase in electricity costs.
Microsoft also states that a server rack with 96 Maia 300 accelerators will provide up to seven times higher token generation performance than the previous generation Maia 200 system.
The new architecture involves the use of several specialized computing units manufactured using the advanced N2P technological process of Taiwanese TSMC.
Each accelerator will have eight blocks of high-speed HBM4 memory with a bandwidth of up to 18 TB/s. A full server rack will contain 27.6 TB of such memory and consume up to 167 kW of electricity.
However, the stated advantages of Maia 300 are Microsoft's own estimates. Their confirmation in large industrial deployments will be possible after the new equipment begins use.
Microsoft already uses its own accelerators
The previous generation of Maia 200 processors already works in Microsoft's infrastructure and serves part of the requests to Copilot.
The equipment is also used by Microsoft's internal AI team for developing new models. By the end of 2026, the company plans to expand its application to GitHub Copilot and the Microsoft Foundry platform.
Currently, Maia 200 is being deployed in data centers in the USA, Europe, and Australia.
According to the corporation, in certain working scenarios these systems already provide up to 1.7 times more tokens per watt compared to leading GPU-based infrastructure.
The new generation of Maia 300 is oriented towards more complex tasks: processing large volumes of text, lengthy model reasoning, operation of autonomous AI agents, and execution of multi-step operations.
This is especially important against the backdrop of the spread of AI assistants capable not only of answering questions but also of independently performing work tasks.
Microsoft does not plan to completely abandon Nvidia
Expansion of its own AI chip production intensifies competition in the market where Nvidia remains one of the main suppliers of equipment for training and running artificial intelligence models.
However, Microsoft has not announced a complete abandonment of Nvidia or AMD accelerators.
As president of the Azure Hardware Systems division Rani Borkar clarified to Axios, the company's strategy involves combining its own processors with third-party manufacturer equipment.
The transition to Maia as the main platform primarily concerns Microsoft's own AI products. It does not mean that all Azure cloud service customers, including OpenAI, will be required to switch to these processors.
In addition, scaling systems to multi-gigawatt capacity will occur gradually. The deployment is scheduled to begin in 2027, but reaching full scale will require more time.
Other technology corporations are also developing their own AI accelerators. Google uses the TPU processor family, Amazon develops Trainium, and Meta also invests in specialized equipment.
For Nvidia, this means additional competition from its largest customers. At the same time, growth in overall demand for computing power may support graphics processor sales even amid the proliferation of alternative solutions.
The main economic goal of Microsoft is to reduce the cost of artificial intelligence operations and serve more users with the same energy resources. As Copilot and autonomous AI agents spread, computing costs become an increasingly important factor in the profitability of technology services.
According to materials: Microsoft, Microsoft Azure, Axios