sequenced.ai
Articles/Data & analytics/Blueprint//8 min read

Where Ciena fits when an AI workload spans two data centers

Ciena supplies optical transport for links between facilities. The right design starts with the traffic that has to cross the fiber and what happens if a route fails.

By Sequenced deskAI-assisted, source-led · how we work
Visit Ciena website ↗
WaveserverTransport platformConnects client traffic to optical wavelengths.
WaveLogicCoherent opticsEncodes and receives data on the optical line.
NavigatorNetwork controlProvides an operations view across network layers.
Metro DCIDeployment useLinks separate facilities in a metropolitan area.
Ciena mark
Cienaciena.com · independent research

Represent this company? Verify your work email to access its workspace, or send the desk a factual correction.

Two buildings can each have space and power for an AI cluster and still make a poor home for one shared workload. A training job that exchanges intermediate results constantly makes different demands on the connection from a job that sends a finished checkpoint once an hour. Ciena supplies optical networking equipment, software and services for connecting facilities. Deciding what actually has to cross the fiber is the first step toward deciding whether to buy its equipment, lease a carrier service or move the work.

In brief
  1. 01What Ciena supplies Optical transport platforms, coherent optics, line systems and software that operators use to manage network services.
  2. 02Who makes the decision Operators with fiber or a leased optical path, and AI infrastructure buyers deciding what service to request from a carrier.
  3. 03What to measure Usable client capacity, round-trip delay, recovery after a path failure and the impact of all three on the intended application.

01 / ProductThe optical line and the port facing the customer are different things

Ciena’s data-center interconnect overview distinguishes campus, metropolitan and longer-distance links. It also identifies scale-across designs for distributed AI training. These labels describe different engineering situations. The fiber length, route diversity and installed equipment influence which signals a path can carry. A buyer with two adjacent buildings and a buyer crossing a metropolitan network can both need more capacity while facing quite different propagation delays and failure scenarios.

Waveserver is a family of transport platforms. At a high level, traffic from client equipment enters at one end and is carried over an optical line to another location. WaveLogic 6 Extreme is coherent optical technology used in platforms including Waveserver. Ciena gives an upper rate of 1.6 terabits per second for a wavelength with this generation. That describes an optical capability. It does not specify the bandwidth delivered to an individual Ethernet port on a particular service, or the achievable rate on any chosen fiber route.

The software layer matters to the operator that has to keep the service running. Navigator Network Control Suite brings visibility and control to the network, including the optical underlay. When an application slows down, a team needs to distinguish a problem with a server, a congested client handoff and a degradation on the optical path. A control system can supply relevant network evidence. It still takes people who can correlate that evidence with an application’s timing.

02 / AudienceWho should own a wavelength and who should buy a service?

An operator that controls dark fiber and staffs optical operations can evaluate the equipment itself. It has choices about where to terminate wavelengths, how to protect them and how to add capacity later. A research company leasing a few racks may have little reason to take those decisions in-house. It can ask its carrier for two handoffs, a capacity commitment, a delay range and an agreed recovery condition. Ciena equipment may sit underneath either arrangement.

There is a real service example behind this distinction. In its September 2026 Auckland announcement, Spark Wholesale described using Ciena’s WaveLogic, Waveserver and Navigator. At announcement it listed 10G, 100G and 400G services and described up to 800G as a future option. The current Spark Wholesale service page now lists all four speeds, including 800 Gbps. It specifies fixed paths without optical restoration and offers multiple fixed paths where available. Confirm which sites and paths the operator can supply. A 1.6 Tb/s wavelength and an 800 Gbps client handoff describe different parts of the service.

03 / WorkflowA proposed interconnect study begins with two clocks

Suppose a model developer trains in one facility and keeps a second copy of recent checkpoints in another. One clock measures how long a checkpoint may take to arrive. The other measures how old the latest usable remote checkpoint may be when a site goes offline. Define both. If training writes every hour and each transfer completes fifty minutes later, the age of the recoverable state varies widely across the hour. A fast optical port alone cannot resolve a mismatch between that schedule and the recovery requirement.

Write a traffic budget using the actual largest checkpoint, any concurrent replication and the period available to transmit it. As an illustrative calculation, moving one decimal terabyte in ten minutes requires about 13.3 gigabits per second of useful payload throughout that interval. Protocol overhead, retries and competing traffic increase the capacity needed at the client port. This estimate is a starting point for a service discussion. It cannot establish which wavelength, transponder or line configuration the route requires.

Next distinguish background transfers from synchronization between training workers. A copy can often tolerate an occasional delay if it still finishes inside its window. Workers waiting for each other at frequent barriers can lose useful compute time on every exchange. Replay the specific communication pattern before distributing a tightly coupled job. It may be more effective to keep a training group within one facility and use the link to distribute datasets or completed models.

Finally test the secondary path during the same workload. Two circuit identifiers can share a building entrance or a length of duct. Ask the operator to explain the diversity it can attest to and what remains shared. Record the application-visible interruption during a planned switchover and how long the remote checkpoint takes to become usable. The resulting study gives the network team and the training team a common answer about what the link is protecting.

04 / PricingBuying transport equipment differs from ordering bandwidth

PurchaseCommercial routeSpecify before comparing
Owned optical systemRequest an engineered Ciena or partner quoteClient and line interfaces, fiber path, spares and support
Carrier-provided DCIOrder service from the network operatorEndpoints, protected capacity, handoff and term
Management software and servicesScope alongside the operated equipmentCovered network, operational access and responsibilities

Public commercial routes consulted 19 November 2026. Ciena sales, Waveserver and the Spark service announcement.

The cited product and sales pages direct an equipment buyer to a sales discussion. They do not publish a universal price per wavelength or per transported terabyte. Ask for the initial system, the cost of adding a second service and the conditions that trigger additional line equipment. A low initial quote that omits the other end of a link, a protected path or ongoing support cannot be compared directly with a carrier’s recurring service charge.

For a carrier proposal, keep construction and lead time visible. A recurring capacity price may assume that both sites are already connected. For an owned system, include the fiber arrangement, operating staff and replacement plan. Then compare the alternatives over the same period and at the same protection level. The meaningful unit is the protected service between named endpoints that meets the workload’s deadline. A price computed from maximum optical line rate can hide the most expensive parts of the design.

05 / DistinctionsA useful comparison follows the data out of the rack

Arista’s networking Blueprint discusses the switching layer closer to servers and their traffic. A fabric inside a building may deliver excellent performance while the link to the other building is constrained. Conversely, adding optical capacity can do little for a job that is waiting on local switching or storage. Measure where the delay first appears. The two suppliers’ product families can address different portions of one end-to-end path.

Corning’s Blueprint provides another boundary to keep in view. The physical fiber infrastructure and its connectivity determine the available path, while coherent systems make use of the optical medium. If the two facilities lack suitable routes or genuinely independent entrances, replacing the electronics may leave the main reliability issue unresolved. Separating decisions about the medium from decisions about the terminal equipment makes competing proposals easier to read.

06 / QuestionsWhat survives maintenance or a loss of one path?

Ask for performance at the capacity the customer will actually receive during maintenance. A protected design may briefly interrupt service, or it may leave less headroom for checkpoint transfers when only one path remains. Name the activity that gets priority in that condition. If a backup copy and an interactive inference service share the interconnect, the effect on the business depends on how the remaining bandwidth is assigned as well as on how quickly the optical network recovers.

Clarify who will see the information needed to investigate. A carrier may have detailed optical telemetry while its customer receives service-level reports. An owner of an optical system may gain more direct visibility but must also interpret the data. Agree on the timestamps, performance measures and escalation contacts both teams will use. This is particularly important when an optical link stays up while application performance deteriorates, because a simple up-or-down service check can miss the incident.

07 / DecisionSelect the service at the level you can operate

01

You control suitable fiber and optical operations

Bring a route survey, a client traffic budget and the failure cases the system must tolerate. Compare designs at a realistic initial load and the next planned expansion.

Request a route-specific engineered proposal.
02

You need connectivity between leased racks

Specify two named endpoints, useful payload capacity, recovery expectations and a delivery date. Let a provider identify the underlying equipment.

Compare service commitments and verified diversity.
03

You want to split a training job across buildings

Measure synchronization behavior on a representative link before deciding where to locate the workers. Test the effect of both steady delay and a path change.

Move the workload only when the application evidence supports it.
What should we explore next?

A business worth understanding.

Suggest your business or one you find interesting. Tell us what you want to understand about its product, positioning, design or workflows.

Suggestions are free. Selection and publication stay with the desk.

Sources
Filed under Data & analyticsCompany CienaNot affiliated with CienaRequest a correctionRequest a refresh by email

Continue reading

All in this category