Qualcomm and AWS Partner on Custom Silicon for AI Datacentre Scaling
The collaboration targets high-speed optical connectivity and custom silicon design to address the growing infrastructure demands of large-scale AI workloads.
Qualcomm Technologies and Amazon Web Services (AWS) entered a multi-generation strategic partnership on September 10, 2026, to develop custom silicon and high-performance connectivity solutions for large-scale artificial intelligence datacentres. The collaboration focuses on optimizing hardware for AI inference and optical interconnects capable of reaching speeds up to 1.6Tbps.
This initiative addresses the rising demand for compute, storage, and memory bandwidth as organizations face the physical limits of existing network architectures. By integrating Qualcomm’s SerDes and digital signal processing (DSP) technologies with AWS infrastructure, the companies aim to improve power efficiency across distributed AI systems.
Qualcomm Incorporated president and CEO Cristiano Amon stated that the partnership leverages decades of expertise in advanced processing to deliver performance gains in power-constrained environments. The design effort prioritizes system-level integration to support the massive scale required by modern generative AI workloads.
Prasad Kalyanaraman, vice-president of AWS, noted that the joint engineering effort reflects a shared commitment to building more cost-effective and performant infrastructure for cloud customers. The collaboration extends beyond hardware production to include the use of Amazon Bedrock for electronic design automation (EDA) workflows.
Utilizing AWS cloud resources for chip design allows Qualcomm to accelerate its internal development cycles and reduce the time-to-market for future silicon iterations. This workflow integration represents a shift toward cloud-native hardware engineering processes within the semiconductor industry.
The agreement follows Qualcomm’s recent acquisition of Modular, a software stack provider that enables AI models to run efficiently across heterogeneous hardware architectures. This acquisition provides a unified platform that supports central processing units, graphics processing units, and custom application-specific integrated circuits without requiring extensive code rewrites.
Integrating Modular’s technology into the broader Qualcomm portfolio aims to simplify the deployment of agentic AI across both datacentre and edge environments. By combining this software foundation with the new custom silicon initiatives, the company seeks to strengthen its position among hyperscalers and enterprise developers.
The technical implementation of 1.6Tbps optical connectivity relies heavily on the integration of advanced SerDes and DSP components to maintain signal integrity over high-density fiber links. These components are essential for reducing the energy-per-bit ratio, a critical metric for datacentres struggling with the thermal constraints of high-bandwidth AI clusters.
Engineers are focusing on the convergence of SerDes lanes to ensure that data movement does not become the primary bottleneck for inference tasks. By optimizing the DSP layer, the partnership aims to provide a more consistent latency profile, which is vital for real-time AI applications that require rapid model response times.
The push for specialized silicon reflects a broader industry trend toward disaggregated compute, where software-defined optimization is increasingly required to maximize the utility of physical accelerators. As AI models grow in complexity, the bottleneck is shifting from raw processing power to the efficiency of data movement and system-level orchestration.
Industry analysts observe that the transition toward 1.6Tbps connectivity is critical for maintaining performance in clusters where network latency frequently limits the throughput of large-scale training and inference tasks. The partnership positions Qualcomm to capture a larger share of the infrastructure market by providing a bridge between high-performance silicon and the software stacks that manage them.
Modular’s software stack addresses the MLOps challenge of heterogeneous hardware fragmentation by providing a unified interface for model execution. This allows developers to deploy models across diverse silicon architectures without the need for manual optimization or platform-specific rewrites, significantly reducing the operational overhead of managing large-scale AI deployments.
Future milestones for the partnership will likely involve the deployment of these custom chips within AWS datacentres and the subsequent optimization of the Modular software stack for these specific architectures. Stakeholders will monitor the integration progress as both firms attempt to mitigate the rising energy costs associated with maintaining massive AI-ready compute clusters.
