Qualcomm is expanding rapidly from its mobile roots into AI data center infrastructure, today announcing with Amazon Web Services that it will co-develop custom AI inference chips and optical digital signal processing (like its Dragonfly portfolio) as a foundation for custom optics solutions. 

The partnership may help chip away at Nvidia’s current dominance in the realm of AI hardware, if it succeeds in offering cost-effective solutions for inference. Its move into cloud infrastructure is helping it jockey for a better position in the data center market – marking a swift pivot from its traditional focus on mobile processors.

With the Qualcomm deal, AWS gets a highly-specialized partner whose decades of building smartphone-driven high performance/low-power system architectures might help it optimize performance-per-watt, and at a lower cost than would be possible with Nvidia. For example, Qualcomm’s recently announced AI200 and AI250 accelerator chips will feature entire liquid-cooled server racks designed specifically for low-cost, low-power AI inference.

In addition, the pair announced they plan to build high-performance 1.6-terabit-per-second optical connectivity solutions based on Qualcomm’s SerDes and optical DSP technology. 

Rather than a standard upfront cash payment, the duo have agreed on a performance-tied equity warrant structure that aligns long-term financial outcomes with commercial milestones. In the deal, Qualcomm issued Amazon financial warrants to acquire up to 25 million shares of Qualcomm common stock. That means the shares will vest in incremental tranches that are released as Amazon fulfills commercial commitments.

This deal, along with the deal Qualcomm landed with Meta in June for data center CPUs shows hyperscalers value the one-two punch of compute + optical-driven networking solutions.