Microsoft will increase its reliance on homegrown artificial intelligence chips starting with the third-generation Maia accelerator, a move designed to reduce dependence on Nvidia graphics processing units while bolstering the company's AI infrastructure. The shift follows years of reliance on Nvidia hardware to handle incoming requests across the software giant's cloud and software services.
The company's Azure cloud remains a central strength, hosting models for clients and powering AI features in Microsoft 365 Copilot and GitHub Copilot. Although Microsoft develops in-house models, they have not yet achieved the market dominance seen with OpenAI, Anthropic, or Google. The push for proprietary silicon stems from sustained demand for advanced processors and a broader industry desire for diverse hardware options. While Nvidia GPUs have been the standard for AI workloads and continue to fill Microsoft's data centers, model developers are increasingly adopting silicon from Amazon, Advanced Micro Devices, Cerebras, Google, and Microsoft.
For general-purpose computing, Microsoft already utilizes its own central processing units alongside hardware from Intel and AMD. Amy Hood, the company's finance chief, stated at a Goldman Sachs conference that the use of proprietary chips has contributed to margin improvement over the past decade. Microsoft introduced its first AI chip, the Maia 100, in 2023 and launched the Maia 200 in January. In July, CEO Satya Nadella told analysts that the Maia 200 delivers 30% better performance per dollar than the latest generation hardware in Microsoft's fleet and is currently handling OpenAI models.
Production of the Maia 300 is scheduled to begin in 2027. Scott Guthrie, Executive Vice President, noted in a blog post that Microsoft is targeting multi-gigawatt deployments of the Maia family. Alistair Speirs, a general manager at Microsoft, declined to provide a specific timeline for these deployments, citing the rapid pace of change in the industry. "We reserve a little bit of optionality there, because this is an industry that's changing very fast," Speirs said.
The Maia 300 is designed to process up to seven times more AI tokens per rack than its predecessor. Microsoft plans to pack 96 Maia 300 chips into each data center rack, double the density of the Maia 200. When compared with GPUs, the new chip offers 40% more tokens per megawatt. Speirs emphasized that new architectures are necessary to improve performance and power efficiency, noting that power usage translates directly into operational costs for data centers. Jim Kleewein, a technical fellow at Microsoft, said parts of Microsoft 365 Copilot should see speed improvements with the Maia 300.
Evercore ISI analysts estimated in a note distributed on Wednesday that Maia chips will account for approximately 10% of Microsoft's AI chip spending by the first quarter of 2027. The firm maintains a buy recommendation on Microsoft shares. Nvidia remains a significant partner for Microsoft; CEO Jensen Huang spoke at an event last week where Microsoft announced a Surface laptop featuring an Nvidia processor. Speirs confirmed that Microsoft continues to partner with both Nvidia and AMD for model training.
Unlike tensor processing units from Google or Inferentia and Trainium chips from Amazon, Microsoft's accelerators are not available for hourly rental by other companies. Patrick Moorhead, CEO of Moor Insights and Strategy, observed that no outside parties have tested the chips yet.