AWS will use custom Qualcomm chips for its inference workloads

Qualcomm will design custom chips for AWS across several product generations, focused on inference workloads where the energy cost per token drives the bottom line. The two companies are also developing optical interconnects with bandwidth of up to 1.6 terabits to carry the growing data traffic inside AI infrastructure. AWS is adding the designs alongside its own Trainium, Graviton and Nitro families, while Qualcomm uses AWS services such as Amazon Bedrock to speed up its chip design work. Amazon is Qualcomm's third large data center win since June: Meta is taking the Dragonfly C1000 server processor and Microsoft is deploying Qualcomm's HBC memory architecture in Azure. Qualcomm targets $15 billion in data center revenue by 2029.