Rob Shore, head of optical networks solution marketing at Nokia Network Infrastructure, said AI is generating significantly more bandwidth demand between locations, including data centers, regions, and cloud environments
In sum – what to know:
Scale-across expands – Large AI clusters are increasingly extending beyond individual data centers, requiring connectivity between physically separate facilities over hundreds or thousands of kilometers.
Fiber limits loom – Optical innovation is approaching the theoretical capacity limits of a single fiber pair, increasing the importance of multi-fiber architectures.
Power reshapes geography – Limited power availability is influencing where AI data centers are built, creating demand for new fiber and long-distance optical connectivity.
The rapid expansion of AI infrastructure is pushing optical networks toward a new architecture in which connectivity between multiple physical data centers becomes an increasingly important part of the AI infrastructure itself.
Rob Shore, head of optical networks solution marketing at Nokia Network Infrastructure, said AI is generating significantly more bandwidth demand between locations, including data centers, regions, and cloud environments. In many cases, that demand is beginning to exceed what can be accommodated through a single fiber pair.
This is accelerating the move toward multi-fiber deployments and changing the emphasis of optical innovation. Rather than focusing exclusively on increasing the amount of capacity carried over an individual fiber, the industry is increasingly looking at how to scale capacity across multiple fibers while reducing cost, power, and space requirements and keeping networks simple to operate.
The shift is also being driven by the growing size of AI compute clusters. Shore said the largest GPU clusters now exceed 500,000 GPUs, with the industry heading toward 1 million-GPU clusters potentially by the end of 2026.
At that scale, the physical and power requirements of a single facility become increasingly difficult to manage. Once GPU clusters extend beyond a single building, connectivity that would previously have been considered external to the data center becomes part of the architecture required to operate the AI environment.
Shore said connectivity that was once primarily required inside the data center for scale-up and scale-out is now spilling into the WAN. He noted that the industry has increasingly adopted the term “scale-across” to describe data center interconnect architectures that connect GPUs or other accelerators across multiple physical data centers.
The distances involved can extend to hundreds or even thousands of kilometers, turning optical transport into an extension of the AI compute environment. This architecture is creating demand for technologies that can deliver high capacity without adding excessive space or power requirements. Shore pointed to coherent pluggables and multi-fiber amplifiers as examples of technologies supporting scale-across deployments.
At the same time, the industry is approaching fundamental limits on how much information can be transmitted over an individual fiber pair. Shore said optical innovation remains important for reducing the cost, space, and power required for data center connectivity, but the laws of physics increasingly constrain the amount of additional capacity that can be extracted from a single pair.
The Nokia executive said hollow-core fiber could eventually increase transmission spectrum and capacity per fiber, but stressed that the technology remains far from wide-scale commercial deployment. In the nearer term, therefore, the focus is increasingly on making multi-fiber networks practical to deploy and operate.
The geographical distribution of AI infrastructure is adding another dimension to the challenge. Power availability is increasingly determining where AI data centers can be located, meaning new facilities may be built in places where sufficient power is available rather than where fiber infrastructure already exists or where facilities are close to population centers. That creates a requirement for new fiber builds and longer-distance connectivity.
“Power availability increasingly dictates where AI DCs built, changing network geography. Facilities built where power accessible versus where fiber exists or near population centers. More geographically distributed footprint requiring new fiber builds and long-distance connectivity,” Shore said.
For network operators and infrastructure suppliers, the resulting expansion represents a significant buildout challenge, but also provides an opportunity to design new networks around current-generation fiber and optical transmission technologies.
Shore said new builds can be optimized around factors including amplifier spacing, physical footprint and energy efficiency. The scale and strategic importance of AI infrastructure is also encouraging closer cooperation between AI and cloud providers and network infrastructure suppliers.
The result is an optical network environment increasingly designed not simply to connect data centers, but to allow distributed AI compute resources to function as part of a larger, geographically dispersed infrastructure.
The interview with Nokia’s Rob Shore is part of a report published by RCR Wireless News and RCRTech, titled Scaling Optical Networks for the Hyperscale and AI Era, which can be accessed by clicking here.
