Qualcomm and AWS Agree to Co-Design Custom Inference Silicon
Qualcomm and Amazon Web Services unveiled a multi-generation collaboration to co-design custom AI-inference chips and optical connectivity rated up to 1.6T. The commercial package includes a warrant allowing Amazon to acquire up to 25 million Qualcomm shares. Qualcomm’s stock jumped about 10% on the news.
Hyperscalers have spent two years designing around NVIDIA supply constraints. Google has TPUs, Amazon has Trainium and Inferentia, Microsoft and Meta have custom efforts, and now AWS is tying a major mobile-and-edge silicon vendor into future inference generations. The optical piece matters as much as the die: moving activations between racks is becoming as scarce as FLOPs.
A warrant for 25 million shares aligns incentives if the chips ship. It also signals that Amazon wants more than a purchase order — it wants a claim on Qualcomm’s upside if inference silicon becomes a second major product line beside smartphones.
This does not dethrone NVIDIA in training. It does add another serious attempt to make serving frontier models cheaper and less GPU-bound, which is where most of the long-run cost of agents will sit.
Key takeaway — AWS is buying a multi-year inference roadmap and an equity hook into Qualcomm. The deal is about serving models at scale, not replacing NVIDIA overnight in training clusters.
Photo: Unsplash (circuit board). Sources: AI Weekly, BitsMinds, Analytics India Magazine coverage of the Qualcomm–Amazon announcement.
