Microsoft is intensifying its push into custom AI silicon, planning to deploy its Maia-series chips at gigawatt scale as it gradually reduces reliance on Nvidia GPUs and seeks to control AI infrastructure costs by boosting inference efficiency.

On October 12, Microsoft announced that its next-generation Maia 300 chip will begin deployment in data centers next year, purpose-built for AI inference workloads. Microsoft expects Maia 300 to process 40% more tokens per megawatt than mainstream GPU systems, meaning it can handle more AI workloads for the same energy consumption.

Unlike Google and Amazon, which have already opened their custom chip compute capacity to cloud customers, Microsoft currently has no intention of renting out Maia compute directly to external clients. This choice reflects Microsoft's priorities: first use its own chips to meet its massive internal AI service demand, then convert hardware efficiency advantages into business competitiveness.

Gigawatt-scale deployment, with energy efficiency as the core rationale

Microsoft's push into custom chips is fundamentally about improving AI inference efficiency and lowering operating costs. Inference refers to using a trained AI model to perform actual tasks, such as summarizing meeting notes or answering research questions, and it represents a major source of compute demand in running AI applications.

Microsoft says the Maia series is specifically designed for inference workloads in its own data centers, enabling it to run relevant tasks more efficiently and reducing the need to purchase Nvidia GPUs. The latter remains the primary source of compute for Microsoft's inference business.

Rani Borkar, who heads Microsoft's data center hardware business, said: "We believe we have built the best product for these workloads. We will deploy Maia at gigawatt scale starting next year, and we are backing that confidence with action."

Previous generation already in place across multiple regions, new model to follow next year

Microsoft's custom chip strategy has been underway for some time. Early this year, the company began installing the previous-generation Maia 200 in data centers, and the chip is now used to power AI software including Copilot as well as other tasks.

According to Borkar, Maia 200 can support OpenAI models, Microsoft's own AI models and several open-source model families, with deployments spanning 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 hopes to leverage its energy efficiency advantages to further optimize compute allocation and reduce inference costs amid its continuously expanding AI business.

In the custom chip race, Microsoft chases Google and Amazon

Against the backdrop of cloud computing giants developing their own AI chips, Microsoft's efforts are relatively late. Google and Amazon started earlier and have established first-mover advantages in custom chips, opening related compute capacity to cloud customers.

Microsoft has chosen a different commercial path. Borkar said: "Our own AI services business is already large enough, and we want to serve ourselves first."

This means that at this stage Microsoft prefers to use Maia chips for internal AI operations rather than generating direct revenue by renting out compute. For investors, Maia's commercial value will depend more on whether it can improve the operating efficiency of AI services, lower inference costs and ultimately improve the profit margins of related businesses, rather than on how much direct revenue a chip-rental business could generate.