Microsoft weighing custom AI chips as it reassesses dependence on Nvidia GPUs

Microsoft is reconsidering how much it leans on Nvidia for AI infrastructure, according to a report published by Yahoo Finance on Monday. The article frames the issue around costs, arguing that custom silicon could help Microsoft reshape the economics of running large language models and other AI workloads at scale.

At the center of the discussion is the role of Nvidia graphics processing units, or GPUs, which have become a dominant building block for training and deploying many AI systems. Nvidia’s chips are widely used across data centers, and the report characterizes Microsoft’s existing approach as meaningfully exposed to that supplier relationship.

The Yahoo Finance piece also points to an alternative strategy: designing or sourcing purpose-built AI hardware, often referred to as “custom silicon.” In practice, custom silicon can mean chips tailored to a company’s software stack and performance targets, potentially lowering per-inference costs and reducing dependency on any single external supplier for key components.

While the report links the “custom silicon” idea to Microsoft’s broader AI infrastructure planning, it does not, based on the information available for this draft, specify the timeline, the exact chip approach, or whether Microsoft is pursuing its own end-to-end chip design or partnering for certain components. Those are the kinds of operational details that would typically determine how quickly any cost benefit could show up in results.

For Nvidia, the prospect of even partial shifts by a major cloud customer is strategically sensitive. Nvidia’s market strength in AI has been tied not only to chip performance, but to the ecosystem around its hardware, including software tools and platforms that help developers move workloads efficiently. If customers diversify hardware paths, Nvidia can still benefit from continued demand, but its share of a given compute supply chain could face incremental pressure.

In the wider technology sector, the “rethink” theme reflects a familiar tension in AI infrastructure: hyperscalers want best-in-class performance and reliability, but they also want leverage on pricing, availability, and long-term supply commitments. Custom silicon is one way large buyers seek more control over their cost structure, especially as inference costs can become a larger share of total spending once models move from experimentation to steady production use.

Still, important questions remain unanswered in the report as presented here. It is not clear what portion of Microsoft’s AI workloads the article implies would move away from Nvidia, nor what performance, reliability, or deployment hurdles Microsoft would need to address to make custom hardware viable at scale. Large cloud providers typically run many workloads with varied latency, throughput, and power requirements, which can complicate a wholesale switch.

Investors and industry watchers will likely focus on any additional indicates from Microsoft about hardware roadmaps, procurement plans, or partnerships tied to custom AI chips. For Nvidia, the near-term watch items would include whether major cloud customers broaden their supplier mix and whether Nvidia continues to see demand strength in the AI accelerator market despite the prospect of more heterogeneous compute.