Advances in optical transport technology are helping operators extract more capacity from existing fiber infrastructure, reckons the analyst firm
In sum – what to know:
Early stage – AI traffic is not yet placing significant pressure on optical networks, but metro, long-haul and subsea infrastructure must be expanded ahead of future AI-driven applications.
Fiber demand – Higher-capacity optical technologies such as 800G, 1.6T and ZR+ improve spectral efficiency, but Dell’Oro says they will complement—not replace—the need for new fiber deployments.
Power bottleneck – While optical transport investment is expected to accelerate, power availability remains the primary constraint for scaling AI infrastructure over the next three to five years.
Artificial intelligence has yet to become a significant driver of optical network utilization, but network operators and infrastructure providers will need to expand metro, long-haul, and subsea infrastructure ahead of future AI-driven applications, according to analysts at Dell’Oro Group.
Speaking with RCR Wireless News, Jimmy Yu, vice president at Dell’Oro Group, said the industry remains in the very early stages of AI adoption from a networking perspective, even though infrastructure must be built to the scale required to support future AI applications.
“We are at the early stage of AI traffic being a driver of network utilization. In fact, I would say that we are at the very early stages of AI. So, I doubt AI traffic is putting any pressure on the optical network at the moment. That said, regardless of whether AI traffic puts pressure on the current network, the infrastructure, including metro, long-haul, and subsea, must be built out to the scale required to enable AI-type applications first,” Yu said.
Yu added that advances in optical transport technology are helping operators extract more capacity from existing fiber infrastructure, but argued that those improvements alone will not eliminate the need for additional fiber deployments.
According to Dell’Oro, technologies such as 1.6 Tbps-capable wavelengths and the use of both C-band and L-band are improving spectral efficiency and helping delay new fiber deployments. However, those gains are approaching fundamental physical limits while AI infrastructure continues to expand.
“The current optical technologies, such as 1.6 Tbps-capable wavelengths and the use of both C-band and L-band, help reduce the need for new fiber deployments by improving spectral efficiencies. However, as we have nearly reached Shannon’s Limit, the benefits are diminishing. For this reason, there is a strong need for new fiber deployments. Adding fuel to this hot demand for fiber is the accelerated construction of new data centers and the interconnection of AI data centers to create larger virtual AI factories. So, while new optical technologies help, I don’t think they will reduce the demand for new fiber deployments,” Yu added.
The need to interconnect larger pools of AI compute servers across multiple facilities is increasing the importance of data center interconnect (DCI) architectures. Yu explained that AI clusters increasingly need to scale across separate data centers because of power limitations at individual sites, creating demand for high-capacity optical interconnect technologies.
“The need for a larger pool of interconnected AI compute servers and the limitation of power resources at a data center site is creating the need to scale-across separate data centers to form a larger virtual AI factory with data center interconnect (DCI). The current architecture being deployed for scale-across DCI is IP-over-DWDM (IPoDWDM), where large quantities of 800 Gbps ZR+ pluggable optics are installed on a data center switch/router and multiplexed into hundreds of fiber pairs between AI data centers,” he said.
While networking technologies continue to evolve, Alex Cordovil, research director at Dell’Oro Group, said power remains the most significant and difficult bottleneck facing AI infrastructure expansion.
“The constraints are interconnected but not equal: fiber and deployment timelines are manageable with investment and execution, while power is the most significant and difficult bottleneck,” said Cordovil.
Looking ahead, Dell’Oro expects power to remain the primary limiting factor for AI infrastructure over the next three to five years. “Power is the primary constraint over the next 3–5 years. While compute supply and network capacity are scaling on clear roadmaps, power depends on slower grid and generation build-outs. Efficiency gains in AI are likely to be offset by rising demand, reinforcing power as the key bottleneck,” Cordovil added.
The interview with Dell’Oro Group’s Jimmy Yu and Alex Cordovil is part of a recent 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.
