Microsoft to replace Nvidia GPUs with its own Maia accelerators

Microsoft CTO Kevin Scott announced the company plans to transition most of its AI workloads from Nvidia and AMD GPUs to its own homegrown Maia accelerators in the coming years. The strategic shift reflects Microsoft’s pursuit of better price-performance ratios and greater control over its datacenter infrastructure, positioning the company to compete more effectively with Amazon and Google, who have been developing custom silicon for years.

What you should know: Microsoft is betting big on its second-generation Maia accelerator to challenge GPU dominance in its datacenters.

• The company successfully moved OpenAI’s GPT-3.5 to its first-generation Maia 100 chips in 2023, freeing up valuable GPU capacity for other workloads.
• A next-generation Maia accelerator is reportedly coming to market in 2025 with significantly improved compute, memory, and interconnect performance compared to the current 800 teraFLOPS Maia 100.

The big picture: Microsoft is joining the custom silicon race later than its cloud competitors but with ambitious goals for market disruption.

• Amazon and Google have deployed tens of thousands of their TPUs and Trainium accelerators over the past few years, securing high-profile wins like Anthropic, an AI company.
• However, these hyperscalers still rely heavily on Nvidia and AMD GPUs because enterprise customers continue demanding them for cloud workloads.

What they’re saying: Scott emphasized that system-level optimization drives Microsoft’s custom chip strategy beyond just raw performance metrics.

• “It’s about the entire system design. It’s the networks and cooling, and you want to be able to have the freedom to make decisions that you need to make in order to really optimize your compute for the workload,” Scott told CNBC.
• When asked directly about using mainly Microsoft silicon in datacenters, Scott responded: “Yeah, absolutely.”

Why this matters: The transition represents a fundamental shift in how hyperscale cloud providers approach AI infrastructure economics.

• Performance per dollar has become the critical metric for cloud providers managing massive AI workloads at scale.
• Custom accelerators allow companies like Microsoft to optimize entire system architectures rather than relying on general-purpose GPU solutions.

Broader context: Microsoft’s silicon ambitions extend beyond AI accelerators to include a comprehensive chip portfolio.

• The company is also developing its own Cobalt CPU and specialized platform security silicon for cryptography and key exchange protection.
• This multi-pronged approach mirrors strategies employed by other tech giants seeking greater control over their hardware destiny.