Amazon Bets on Hollow Core Fiber to Keep a Million Chips in Sync

AWS’s push into hollow core fiber (HCF) technology points to a critical bottleneck in scaling AI infrastructure that extends beyond chips. While GPUs and high-bandwidth memory often dominate discussions, AWS principal network development engineer Stephen Callaghan states that a problem in any part of the network can halt an entire GPU operation in systems handling over a million chips. HCF, by allowing light to travel faster through an air-filled core, reduces latency by 30% and extends network reach by 50%. This is particularly relevant for Asia, where data center growth is accelerating, and the physical distances between facilities, such as the new subsea cable linking the western United States and Japan targeted for 2029, demand advanced low-latency solutions. The adoption of HCF by hyperscalers like AWS and Nvidia suggests a new front in the AI infrastructure race. The technology's ability to maintain high-speed data transfer over longer distances directly supports the increasing decentralization of data centers, driven by constraints on power, cooling, and land. For Asian markets, this means that future data center investments will likely prioritize network architecture and advanced fiber optics as much as raw compute power, potentially spurring local innovation in optical fiber manufacturing despite current hurdles like high cost and complex production processes. This development underscores that the competitive edge in AI infrastructure will increasingly depend on optimizing every layer of the stack, from silicon to network fabric. Companies in Asia looking to build or host large-scale AI operations must consider these network advancements to ensure their systems can handle the immense data synchronization requirements of next-generation AI models.