Amazon Web Services on Tuesday said it has tapped Qualcomm to develop customized silicon for AI inference and 1.6 Tbps optical networking gear for its datacenters, in a multi-generation partnership that gives the mobile-chip specialist a foothold inside the largest U.S. cloud provider and adds another supplier to Amazon’s in-house AI chip roadmap.
The two companies disclosed the arrangement in a joint statement without detailing timing, dollar value or the specific products that will be built. Qualcomm said it was bringing decades of leadership in advanced processing and power-efficient compute to the collaboration, and that the aim was to enable the next generation of AI infrastructure. Amazon, which already builds its own Trainium and Inferentia accelerators and uses custom versions of Intel Xeon and AMD EPYC CPUs in its fleet, did not say whether the Qualcomm silicon would be a co-designed accelerator, a chiplet, or a customized version of Qualcomm’s existing datacenter product line.
Qualcomm has been pushing into the datacenter market this year with an AI250-series of Dragonfly rack systems that it announced earlier in the summer. Those systems swap out costly high-bandwidth memory in favor of what the company calls high-bandwidth compute — an architecture that uses cheaper LPDDR memory but shifts some of the compute work into the memory’s base die, which Qualcomm argues delivers higher effective memory bandwidth for AI inference workloads at a lower per-rack cost. The company is also developing a new datacenter CPU that it says will reach 5 GHz and scale to more than 250 cores when it arrives in the second half of 2028.
The networking piece of the AWS deal is more concrete. Amazon said it will use Qualcomm’s serializer-deserializer chips and optical digital signal processors — technology Qualcomm acquired late last year — to build out ports at 1.6 Tbps, doubling the current top-end speeds inside its AI clusters. High-speed optical interconnects have become a major supply-chain bottleneck for the industry as training and inference workloads scale from single racks to clusters of hundreds or thousands of accelerators across multiple racks. The arrangement points to AWS expanding its supply chain to include Qualcomm as a source of optical pluggables and potentially near- or co-packaged optics.
The relationship also runs in the other direction: Qualcomm said it plans to expand its use of Amazon’s Bedrock generative-AI platform for its electronic-design-automation workloads, using AWS models to help it design future chips. Neither company has responded to requests for additional detail on the scope of the multi-generation agreement, though industry analysts said the announcement extends a pattern in which cloud providers layer multiple silicon partners onto their AI roadmaps to hedge against supply constraints from any single vendor.
