Qualcomm is designing custom silicon for AWS across several product generations, aimed at AI inference, and it is designing that silicon on AWS, using Amazon Bedrock among other services to speed its own chip development. The two companies are also building optical interconnects rated at up to 1.6 terabits to move the data volumes piling up inside AI infrastructure. With that, Amazon becomes Qualcomm's third large data-center customer since June.
The other two arrived quickly. Meta picked up the Dragonfly C1000 server processor. Microsoft is deploying Qualcomm's HBC memory architecture in Azure. Three of the largest buyers of AI compute have now placed Qualcomm parts inside their roadmaps in roughly a quarter, which is a different company from the one that spent the past decade being described as a phone chip vendor. Qualcomm's stated target is $15 billion in data-center revenue by 2029.
On the AWS side, the Qualcomm work joins the in-house Trainium, Graviton and Nitro families. The workload in question is inference, running models that are already trained, where the energy cost of each token lands directly on the margin.
That framing is the entire pitch, and it is a real one. Qualcomm's advantage is not peak throughput, it is two decades of designing chips that had to justify every milliwatt against a battery. Inference at scale turns out to be the same optimisation problem with the constraint renamed: the meter is the electricity bill rather than the phone in someone's pocket. A company that never had the option of brute force is an unusually good fit for a workload where brute force is what hurts.
The Bedrock detail is the part worth sitting with. A chip vendor is using a cloud provider's AI services to build the chips that will run that cloud provider's AI services. As engineering, it is unremarkable; EDA workloads have lived in the cloud for years. As positioning, it is a closed loop AWS gets to point at, and it is cleaner marketing than anything either company could have written about latency.
Notably absent from the announcement: any dollar value, any volume, any ship date, and any word on which generations land when. "Several product generations" is a commitment shape, not a schedule. The $15 billion figure is the only number with a year attached to it, and it describes Qualcomm's ambition rather than anything Amazon has agreed to pay.
The other silence is about Trainium. AWS has spent years and a great deal of capital building its own inference and training silicon precisely so it would not have to buy someone else's. Bringing in an outside designer for inference is either a hedge on that program or an expansion of it, and the announcement does not say which. My read is that it is closer to a hedge than AWS would like it to sound: you do not hand multiple product generations of your highest-volume workload to a partner if your own roadmap is covering the demand.
Energy efficiency remains Qualcomm's theme at the other end of the scale too. In March the company's research arm released a framework that runs reasoning models on smartphones.
Which sets up the real test. Qualcomm is betting that inference economics, not training benchmarks, decide who buys silicon over the next four years, and it has three customers who agree enough to try it. If tokens get cheap enough that energy per token stops being the constraint, that $15 billion target loses the argument underneath it.