0
According to Helen Xenos, senior director of portfolio marketing at Ciena, selecting the appropriate optical technology depends on the customer’s primary constraint
In sum – what to know:
AI reshapes DCI – Helen Xenos said AI scale-across deployments require roughly 10 times the capacity of traditional metro data center interconnect while demanding highly reliable connectivity.
Different priorities – Technology choices depend on customer constraints such as power efficiency, fiber availability, spectral efficiency, and deployment speed, rather than a one-size-fits-all architecture.
Many architectures – The Ciena executive argued that AI networking will rely on a coexistence of coherent pluggables, performance transponders, full-spectrum systems, co-packaged optics, and liquid cooling approaches.
Artificial intelligence workloads are reshaping data center interconnect (DCI) networks, but the result will not be a single dominant optical architecture. Instead, operators and hyperscalers are expected to deploy different technologies depending on their infrastructure constraints and business priorities.
That was the key message from Helen Xenos, senior director of portfolio marketing at Ciena, during a recent RCRTech webinar on optical networking and AI infrastructure.
Xenos said data center interconnect now spans a broad range of applications, from metro networks linking facilities within 100 kilometers to backbone and submarine connections, campus deployments, and emerging “scale-across” AI architectures that connect multiple data center sites.
“The big difference with scale-across is massive capacity, 10 times more capacity than what we would see in Metro DCI,” she said. “And also it’s lossless connectivity. So very reliable connectivity is also needed there.”
According to Xenos, selecting the appropriate optical technology depends on the customer’s primary constraint. Some prioritize power efficiency, while others focus on maximizing capacity in limited space, improving spectral efficiency where fiber is scarce, or accelerating deployment.
“It really depends on what is the biggest constraint our customers are facing or what is their top requirement,” she said.
The emergence of AI scale-out networking is already influencing deployments. Xenos noted that Ciena is “already shipping in volume systems to satisfy scale-across deployments,” with innovation occurring simultaneously in coherent optics and photonic line systems.
The Ciena executive explained that combined C- and L-band solutions can increase fiber utilization for certain deployments. “The fact that we have both C and L band variants allows us to double the fiber capacity,” she said.
However, extending AI networks over longer distances presents additional challenges. Using conventional photonic line system designs for large-scale AI infrastructure could require a significant increase in supporting facilities.
“When we’re talking about this new scale-across or AI infrastructure scale—20 petabits per second, as an example—if we’re using conventional line system designs, we would need 22 huts to actually support this capacity,” Xenos said, adding that such an approach “is not a viable solution.”
To address those challenges, she highlighted multi-rail photonic architectures designed to increase density while reducing physical infrastructure requirements.
Deployment speed is becoming another critical consideration as hyperscalers expand AI infrastructure. Xenos noted that conventional optical deployments activate wavelengths individually, creating operational complexity when provisioning large numbers of fibers.
“This is where full spectrum transponders come in that allow you to light up an entire fiber at once,” she said, enabling “cookie cutter deployments” that simplify large-scale rollouts.
Looking ahead, Xenos argued that increasing switch capacity is creating new thermal and density constraints for pluggable optics. She said the industry is pursuing two parallel approaches: co-packaged optics, which places optics closer to switching silicon, and liquid cooling, which enables higher-capacity pluggables while requiring supporting cooling infrastructure.
Summarizing the broader trend, Xenos concluded that AI infrastructure requirements will encourage a mix of optical networking approaches rather than standardization around a single design. “As optical networks evolve for AI, we are going to see the coexistence of many architectures, not just one,” she said.
To access the full webinar, click here.
