Microsoft is poised to increase its utilization of proprietary artificial intelligence chips following years of dependency on Nvidia's graphics processing units (GPUs), as it aims to enhance its AI infrastructure. The company introduced its third-generation Maia AI accelerator on Monday, signaling a shift toward leveraging its own silicon for managing incoming requests.
Microsoft's Azure cloud platform stands as a significant asset in its AI strategy, hosting various models for clients. Tools like Microsoft 365 Copilot and GitHub Copilot deliver AI-enhanced solutions to software developers and office employees. Although the company has developed internal models, they have yet to achieve prominence in a landscape dominated by competitors such as OpenAI, Anthropic, and Google.
The increasing demand for sophisticated processors and the market's desire for diverse options have opened new avenues in chip development. Nvidia's GPUs have long been the gold standard for AI workloads and populate Microsoft's current data centers, where they also support model training for OpenAI and Anthropic.
As market demand surges, AI model leaders are diversifying their silicon sources, incorporating offerings from companies like Amazon, Advanced Micro Devices (AMD), Cerebras, Google, and more recently, Microsoft.
For general computing tasks, Microsoft already employs its own central processing units alongside chips from Intel and AMD. Finance chief Amy Hood remarked at a Goldman Sachs conference last month that the adoption of its own chips has contributed to "margin improvement over the course of a decade."
In 2023, Microsoft launched its first AI chip, the Maia 100, followed by the Maia 200 earlier this year. CEO Satya Nadella mentioned in July that the Maia 200 provides "30% better performance per dollar than the latest generation hardware in our fleet" and that it effectively runs OpenAI models.
Production for the Maia 300 is set to commence in 2027, with ambitious plans for multi-gigawatt deployments within the Maia family, as highlighted in a blog by Executive Vice President Scott Guthrie. However, General Manager Alistair Speirs indicated that a detailed timeline remains undecided.
"We keep some flexibility there since this is a rapidly evolving field, and we need to deploy the most suitable silicon for our models at the right moment," Speirs noted.
Microsoft aims to install 96 Maia 300 chips in each data center rack—double the capacity compared to the Maia 200. The new version is expected to process up to seven times more AI tokens per rack and achieve 40% greater token efficiency per megawatt when compared to traditional GPUs.
"We require new architectures to enhance performance and optimize power usage," Speirs explained. "Power efficiency directly impacts operational costs for data centers."
Jim Kleewein, a technical fellow at Microsoft, mentioned that various components of Microsoft 365 Copilot should see performance improvements with the Maia 300.
In a client update on Wednesday, analysts from Evercore ISI projected that the Maia chips will account for roughly 10% of Microsoft's AI chip expenditures by the first quarter of 2027 and recommended purchasing Microsoft stock.
Despite this push towards self-reliance, Nvidia remains a crucial player in Microsoft’s strategy. CEO Jensen Huang recently spoke at an event where Microsoft revealed a Surface laptop featuring an Nvidia processor. Furthermore, when it comes to model training, Microsoft continues to collaborate with Nvidia and AMD, according to Speirs.
In contrast to Google's Tensor Processing Units or Amazon's Inferentia and Trainium, Microsoft’s accelerators are not available for hourly rental by external companies.
"To date, they have yet to be tested outside of Microsoft," remarked Patrick Moorhead, CEO of advisory firm Moor Insights and Strategy.
