These pipes will need to include metro, regional and long haul links to address all the various AI deployment models. In its 2025 ‘State of the Network’ report, Telegeography found that the transfer of data from training clusters to inference clusters can take multiple forms. Some models are trained in remote locations to get easier and cheaper access to power. These models will have a greater need for haul bandwidth to transfer data. Others locate the training and inference clusters closer together which will have a bigger impact on metro bandwidth demand. In both cases, the size of the data sets and how often they are transferred will have significant impact. The deciding factor which determines if the impact is greater in the metro or long haul network is how close the inference clusters are to the end user.
For these applications, optical connectivity is the only solution Last August, wholesale provider Lumen announced that it was securing 10% of Corning’s global fiber capacity over the next two years just to interconnect AI-enabled data centers. While Corning doesn’t publish how much it produces annually, it’s a lot to the point that even 10% of whatever number that is is still a lot—millions of kilometers of fiber cable.
Not only is more fiber being dedicated to data center connectivity, the service speeds carried over that fiber are increasing as well. 100G wavelength services have replaced 10G as the default connection speed and are now available everywhere while 400G upgrades continue to escalate rapidly. Market research firm Vertical Systems Group reported that data center connectivity among hyperscalers and Cloud providers—and even some large Enterprises—is driving increasing demand for 400 Gbps wavelength services with 800 Gbps services to follow by the end of 2025. Meanwhile, Frost & Sullivan forecasts that North American wavelength services revenue will grow to nearly $6.9 billion in 2028 for $5.03 billion in 2023, a compound annual growth rate of 6.5%, a healthy growth rate for a relatively mature market.
Wholesale communications providers are uniquely poised to take advantage of this opportunity. With their widescale networks capable of offering both long haul and metro links and expertise at deploying and managing wavelength services, they are in the best position to provide the capacity needed to connect these AI clusters. Even for hyperscalers, growth is outpacing their ability to expand the network to every location it needs to reach and they will need to augment their own networks using managed wavelengths on fiber already in the ground. For smaller data center players, carrier-neutral colos and large enterprises, wholesale connectivity will be critical.
To meet this demand, more and more wholesale providers are turning to ‘Spectrum-as-a-Service’ as the model to maximize service and revenue potential in this market. Spectrum-as-a-Service is the partitioning of the usable spectrum among multiple customers, each one getting dedicated capacity that it can use as it chooses so it can expand and connect faster and at its own pace. This capability allows wholesalers to manage and sell large amounts of bandwidth more efficiently by allocating any arbitrary amount of spectrum a customer requests. However, it depends on a flexible transport network built on an open line system with advanced ROADM and NMS capabilities to simplify spectrum management and rapidly provision capacity on demand. With the AI market set to take off, the time is now for wholesale providers to modernize their infrastructure before the opportunity passes them by.
