Microsoft is increasing its investment in self-developed AI chips, planning to deploy the Maia series of chips on a gigawatt scale, gradually reducing its dependence on NVIDIA GPU, and controlling the costs of AI infrastructure by improving inference efficiency.
On October 12th, according to Bloomberg, Microsoft announced on Monday that a new generation of Maia 300 chips will begin to be deployed in data centers next year, specifically designed for AI inference tasks. Microsoft expects that the Maia 300 chips will be able to process token quantities per megawatt 40% higher than mainstream GPU systems, which means they will be capable of handling more AI workloads with the same level of energy consumption.
Unlike Google and Amazon, which have made their self-developed chip computing power available to cloud customers, Microsoft currently has no intention of renting out the computing power of its Maia chips to external customers. This decision reflects Microsoft's priorities: to first use its self-developed chips to meet its own massive AI service needs, and then to transform the hardware efficiency advantages into business competitiveness.
Gigawatt-level deployment; energy efficiency advantages are the core logic.
Microsoft's increased investment in self-developed chips is essentially aimed at improving the inference efficiency of AI and reducing operating costs.
Inference refers to the use of a trained AI model to perform actual tasks, such as summarizing meeting notes or answering research questions. It is also an important source of computing power demand during the operation of AI applications. Microsoft states that the Maia series is specifically designed for inference workloads in their own data centers, enabling more efficient execution of these tasks and reducing the need to purchase NVIDIA GPU. The latter remains the main source of computing power for Microsoft's inference business at present.
Rani Borkar, the person in charge of Microsoft's data center hardware business, stated: "We believe that we have created the best products for these workloads. Starting next year, we will deploy Maia at the gigawatt level to substantiate this confidence with concrete actions."
The previous generation of products has been implemented in many places, and the new model will take over next year.
Microsoft's in-house chip strategy was launched a long time ago. At the beginning of this year, the company began installing the previous generation product Maia 200 in data centers. Currently, this chip is used to power Copilot and other AI software, as well as for other tasks.
Borkar introduces that Maia 200 can support the OpenAI model, Microsoft's self-developed AI model, as well as multiple open-source model series. Its deployment scope covers data centers in the United States, Europe, and Australia, providing AI capability support for products such as Microsoft Teams and Copilot.
The new generation of Maia 300 is expected to be deployed next year. Microsoft hopes to leverage its energy efficiency advantages to further optimize the computing power configuration in its continuously expanding AI business and reduce inference costs.
In the field of self-developed chips, Microsoft is catching up with Google and Amazon.
Against the backdrop of cloud computing giants developing their own AI chips, Microsoft's approach has been relatively lagging behind. Google and Amazon started earlier and have established a first-mover advantage in the field of self-developed chips, offering their computational power to cloud customers for use.
Microsoft, on the other hand, chose a different business path. Borkar states: "Our own AI service business is already large enough, and we prefer to serve ourselves first."
This means that, at present, Microsoft is more inclined to use the Maia chips for its internal AI operations, rather than generating revenue directly by renting out computing power. For investors, the commercial value of Maia will depend more on its ability to improve the operational efficiency of AI services, reduce inference costs, and ultimately enhance the profit margins of related businesses, rather than on how much direct revenue the chip leasing business can generate.
