Ciena Says AI Data Centers Will Need More Than One Optical Network Design

AI DCI Is Splitting Into Several Network Designs

Artificial intelligence buildouts are changing the demands placed on data center interconnect networks, but Ciena is not arguing for a single optical blueprint.

Helen Xenos, the company's senior director of portfolio marketing, said operators and hyperscalers will choose architectures according to the constraint they face first: power, fiber availability, space, spectral efficiency or deployment speed.

That makes the story more specific than a general AI-infrastructure upgrade cycle.

Data center interconnect now spans metro links between facilities within 100 kilometers, backbone and submarine routes, campus networks and newer scale-across designs that connect several data center sites for AI workloads.

The affected market is the transport layer behind AI clusters, not only the servers inside them.

When an AI system is spread across facilities, the interconnect must carry heavy traffic while preserving reliability.

That is why the optical decision moves into the same planning conversation as site layout, fiber supply, power and rollout timing.

Capacity Is Only One Constraint

The clearest technical pressure is scale.

Xenos said scale-across AI deployments can require 10 times the capacity associated with traditional Metro DCI, while also needing lossless and highly reliable connectivity.

That combination narrows the range of optical choices that can work at large AI sites.

Ciena says it is already shipping volume systems for scale-across deployments.

The company is also working across coherent optics and photonic line systems, rather than treating one layer of the network as the only answer.

Combined C- and L-band options are part of that approach because Ciena says the variants can double fiber capacity in some deployments.

That matters for operators with limited fiber routes.

If more capacity can be carried on the same fiber plant, a project may avoid part of the civil-engineering burden that normally comes with new long-haul or metro buildouts.

Ciena's framing is not that every deployment can use the same mix; the choice is constraint-specific.

Long Routes Create A Physical Infrastructure Problem

The harder issue is what happens when AI networking extends beyond a compact campus.

Xenos used 20 petabits per second as an example capacity target for scale-across or AI infrastructure.

With conventional line-system designs, she said that capacity would require 22 huts, a footprint she described as not viable.

That is why Ciena is pointing to multi-rail photonic architectures.

The goal is to raise density while reducing the supporting facilities required to carry very large AI traffic loads over distance.

For cloud operators, the practical question is not only how much bandwidth can be lit, but how many sites, shelters and operational steps are needed to keep that bandwidth usable.

Speed And Heat Shape The Next Choices

Deployment speed is becoming a separate bottleneck.

Conventional optical rollouts activate wavelengths individually, which adds complexity when a project involves many fibers.

Full-spectrum transponders address that problem by lighting an entire fiber at once, giving large deployments a more repeatable rollout pattern.

Switch capacity is pushing another decision point inside the data center.

Xenos identified two parallel paths.

One is co-packaged optics, where optical components sit nearer to the switching silicon.

The other is liquid cooling, which can help operators use higher-capacity pluggables when the site has the supporting cooling systems.

Those choices show why the AI networking debate is moving beyond simple bandwidth expansion.

A project with scarce fiber may lean toward spectral efficiency.

A project under power pressure may make a different optical trade-off.

A project racing to bring new AI capacity online may value full-fiber activation and repeatable deployment steps.

The next checkpoint is whether hyperscale AI projects standardize around a narrow set of optical designs or keep using several architectures side by side, as Ciena expects.