AFL: Why AI infrastructure planning is changing

Author: Joe Peck

In recent years, AI infrastructure discussions centred predominantly on training clusters. Industry attention focused on larger models, sizeable GPU estates, dense scale-out fabrics, and the synchronisation demands created by collective communication across thousands of accelerators. In 2026, however, deployment patterns point to inference as the dominant operational AI workload.

This transition introduces infrastructure behaviours that extend beyond the assumptions of traditional training environments. While much of the industry conversation still focuses on accelerators and compute scale, less attention is given to the implications for network architecture, optical connectivity, and physical infrastructure design.

In response, AFL, a manufacturer of fibre optic cables and connectivity equipment, has developed a whitepaper series to help address that gap.

The first paper, Architecting AI at Scale: From Training Clusters to Inference-Driven Infrastructure, introduces six workload categories representing the evolving AI deployment landscape. These include synchronous training fabrics, throughput inference systems, disaggregated reasoning architectures, heterogeneous decode environments, context-centric infrastructure, and workflow orchestration platforms.

The paper provides practical insight into evolving network behaviours, optical requirements, and multi-domain infrastructure planning.

Future instalments will examine the engineering implications in greater depth. Click here to register to receive email notifications as soon as each paper in the series becomes available.

For more from AFL, click here.