Microsoft plans to deploy its next-generation Maia 300 chips at gigawatt scale next year, with an expected 40% higher token output per megawatt compared to mainstream GPU systems. Microsoft will prioritize using its in-house chips to support internal AI services rather than leasing out computing power, thereby reducing its reliance on NVIDIA GPUs and lowering inference costs to improve the profitability of its AI business.
Microsoft is increasing its investment in proprietary AI chips, planning to deploy the Maia series at gigawatt scale to gradually reduce reliance on NVIDIA GPUs and control AI infrastructure costs by improving inference efficiency.
On October 12, Bloomberg reported that Microsoft announced on Monday that its next-generation Maia 300 chip will begin deployment in data centers next year, specifically designed for AI inference tasks. Microsoft expects the Maia 300 to process 40% more tokens per megawatt than mainstream GPU systems, enabling it to handle greater AI workloads with the same energy consumption.
Unlike Google and Amazon, which have opened their custom chip computing power to cloud customers, Microsoft currently has no plans to directly lease Maia chip computing power to external customers. This decision reflects Microsoft’s priority: first using its custom chips to meet its own massive AI service demands, then converting hardware efficiency advantages into business competitiveness.
Gigawatt-scale deployment, with energy efficiency as the core rationale
Microsoft is investing further in its own chip development to improve AI inference efficiency and reduce operational costs.
Inference refers to using trained AI models to perform real-world tasks, such as summarizing meeting transcripts or answering research questions, and is a major source of computational demand in AI applications. Microsoft states that the Maia series is specifically designed for inference workloads in its own data centers, enabling more efficient execution of these tasks and reducing the need to purchase NVIDIA GPUs, which currently remain the primary source of computational power for Microsoft’s inference operations.
Rani Borkar, Head of Microsoft’s Data Center Hardware Business, said: “We believe we’ve built the best product for these workloads. Starting next year, we will deploy Maia at gigawatt scale to back up this confidence.”
The previous generation has been successfully deployed in multiple regions, and the new model will follow next year.
Microsoft's in-house chip strategy has been underway for some time. Earlier this year, the company began deploying its previous-generation Maia 200 chip in data centers, and it is now powering AI software such as Copilot and other workloads.
Borkar explained that the Maia 200 supports OpenAI models, Microsoft's proprietary AI models, and multiple open-source model families, with deployment across data centers in the United States, Europe, and Australia, providing AI capabilities for products such as Microsoft Teams and Copilot.
The next-generation Maia 300 is expected to begin deployment next year. Microsoft aims to leverage its energy efficiency advantages to further optimize computing resource allocation and reduce inference costs in its continuously expanding AI business.
The in-house chip赛道, Microsoft is catching up with Google and Amazon
Amid the major cloud providers developing their own AI chips, Microsoft’s strategy has been relatively delayed. Google and Amazon entered the market earlier and have established a first-mover advantage in custom chip development, making their related computing power available to cloud customers.
Microsoft has chosen a different business path. "Our own AI services business is already large enough that we want to prioritize serving ourselves," said Borkar.
This means that, at this stage, Microsoft is more inclined to use the Maia chip for its internal AI operations rather than generating direct revenue through leasing 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 profitability of related businesses, rather than on how much direct income the chip leasing business generates.