Qualcomm pushes into AI data centers with Modular and Meta

Qualcomm is moving deeper into AI infrastructure with a package of data center announcements centered on chips, memory, software, and developer distribution. The company introduced Dragonfly as its AI data center brand, including the Dragonfly C1000 CPU, Dragonfly AI300 inference accelerator, and High Bandwidth Compute memory, all aimed at running AI workloads more efficiently while reducing power use and costs.

The acquisition of Modular could be Qualcomm’s most consequential move. Modular’s software lets developers write AI software once and run it across different kinds of chips, including Qualcomm processors and potentially hardware from AMD, Intel, and others. Qualcomm described the goal as a “multi-silicon token factory,” positioning the software layer as a way for AI companies to reduce dependence on Nvidia’s CUDA stack.

Qualcomm also signed Meta as its first customer through a multi-generation deal in which the Dragonfly C1000 CPU will power Meta’s next-generation servers, with production starting in the second half of 2028. A separate agreement with Hugging Face will let its 16 million developers deploy more than over 3 million open AI models on Qualcomm chips across phones, laptops, cars, and AI data centers.

Nvidia is estimated to hold about 80% to 90% of the AI GPU market used to train and run AI models, and reportedly has a 12-month backlog of chip orders. Qualcomm’s strategy could add supply, support mixed hardware deployments, and put pressure on AI compute costs over time.