AWS and Qualcomm strike deal for custom AI silicon
Amazon Web Services has struck a multi-year deal with Qualcomm to develop customised silicon for AI inference workloads and 1.6 Tbps optical networking equipment, though both companies have released scant technical detail. The agreement, which could see AWS commit up to $60 billion in payments through to September 2036, marks Qualcomm's push into becoming a datacentre chip supplier and mirrors similar custom-silicon arrangements AWS already has with Intel and that Microsoft has with AMD.
On the compute side, the collaboration is expected to draw on Qualcomm's forthcoming AI250-series "Dragonfly" rack systems, which replace costly high-bandwidth memory with cheaper LPDDR memory paired with in-memory compute to ease inference bottlenecks; a new datacentre CPU, the C1000, is also due in the second half of 2028. The networking element is clearer: Qualcomm will supply serialiser-deserialisers and optical digital signal processors, likely stemming from its acquisition of Alphawave Semi, to push AWS's port speeds to 1.6 Tbps and address growing optical interconnect demands as AI clusters scale. In return, Qualcomm plans to expand its own use of Amazon's Bedrock platform for chip-design work.
In short
- AWS and Qualcomm sign custom AI chip and networking deal
- Deal could total $60 billion in payments through 2036
- Networking side aims for 1.6 Tbps optical port speeds
Steelman: both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
The deal could give AWS more choice in a market where demand for AI inference and faster data-centre links is rising, while Qualcomm gains a route into a large customer and a new source of growth. If the customised chips and optical components deliver as intended, they may help AWS manage costs and scale AI services; Qualcomm’s design expertise and AWS’s cloud and chip-design tools could also make the partnership productive for both.
The case against
A commitment potentially reaching $60 billion over many years carries substantial execution and commercial risk, especially because the companies have disclosed little technical detail and the proposed compute systems are not yet established products. AWS may face integration challenges and continued reliance on a supplier as its needs evolve, while Qualcomm must prove it can compete as a data-centre chip provider; the deal’s scale could also make it harder to redirect spending if the technology underperforms.
Explained simply
AWS is Amazon’s cloud-computing business: it rents computing power and storage to companies and developers. Qualcomm is best known for making chips for mobile phones, but it is also developing processors and networking equipment for data centres, the large facilities that run online services and AI systems.
AI inference is the stage when a trained AI system responds to a request, such as generating text. Data centres need chips to handle that work and fast connections to move information between machines; as AI systems grow, both can become costly or limit performance. Companies including AWS and Microsoft have worked with chipmakers on custom hardware for their own services.
The agreement also links AWS’s cloud business with Qualcomm’s chip-design work. Its significance depends on how the planned hardware performs and how widely it is used, details that can help explain the companies’ interest in the partnership.
Read the full article at the source →
Originally published by The Register as “Amazon ropes Qualcomm into something, something AI, networking chips”.