Connecting Distributed Data Centers: What's Changing in Optical Transport

AI training clusters, inference fabrics and active-active application architectures have pushed traffic between geographically distributed data centers to levels most WAN transport layers weren't sized for. A single GPU cluster spanning two facilities can now demand more bandwidth between sites than an entire regional network needed a few years ago, and that's forcing operators who thought their data center interconnect (DCI) architecture was settled to take another look. This isn't about connecting racks inside the same building. It's about the fiber, wavelengths and optical line systems that stretch between facilities, sometimes across a metro ring, sometimes across a region.

Power availability sets the distance

A contributing factor to the expanding distances between sites is simply how and why sites are chosen. Facilities were once selected based on their proximity to existing data centers, WAN aggregation points and cloud on-ramps. However, the availability of power has become the dominant selection criterion, and the locations with megawatts to spare rarely sit close to the facilities to which they need to connect. The result is longer interconnect spans by default, thus why amplification and full optical line systems are reappearing on routes that a single unamplified coherent span would once have covered. The optical layer, in effect, absorbs an 'upstream' real estate decision.

Physical path diversity as a site selection variable deserves the same scrutiny. Two facilities that appear separated on a map can still share a conduit, an easement or a single point of failure in the fiber plant. As such, route diversity has to be verified rather than assumed. After all, a protected wavelength is only as reliable as the physical separation of its two paths.

That combination, more bandwidth needed over more distance, is putting new pressure on how operators build the transport layer. Three approaches are competing for that space today: converged IP and optical (IPoDWDM), traditional open line systems (OLS) paired with muxponders or transponders and a fast-growing category of pluggable coherent optics that blurs the line between the two. Here's where each stands and what's worth watching next.

Converged IP and optical keeps gaining ground

IPoDWDM (IP over DWDM) takes the coherent DWDM optic that once lived in a chassis-based transponder and plugs it directly into a router or switch. Instead of a router handing traffic to a separate transponder shelf, which then hands it to a dedicated dense wavelength division multiplexing (DWDM) line system, the router generates the DWDM wavelength itself. That collapses three layers of equipment (IP, optical transport network and DWDM) into one.

For DCI specifically, the appeal is straightforward: fewer boxes between sites, less rack space and power at each facility and one platform to provision and troubleshoot instead of three. WWT's take on converged optical and packet architectures breaks this down into deployment scenarios, from a simple router-to-router link riding a dark fiber pair, up to router-to-ROADM designs that still lean on reconfigurable optical add/drop multiplexers (ROADMs) for passthrough traffic at sites that don't need to terminate every wavelength.

The trend has only accelerated. According to Lightwave, scale-across DCI is now a primary driver of IPoDWDM demand, with hyperscalers using 800ZR+ pluggables to interconnect distributed AI compute clusters. Cignal AI expects 800G coherent pluggable shipments to top $1 billion in revenue in 2026, and a meaningful share of that is DCI, not intra-data-center traffic.

Traditional OLS and muxponder-transponder designs still earn their place

Converged IP and optical isn't the right fit everywhere, and that's where traditional optical line systems earn their place. In this model, transponders or muxponders handle the client-to-line conversion and wavelength multiplexing, while a separate OLS, made up of amplifiers, filters and often ROADMs, handles the actual transport across the fiber. The two layers are managed and scaled independently, which matters for DCI networks with dense wavelength counts, many intermediate sites or a mixed-vendor fiber plant that predates any router-based transport strategy.

We've written about the tradeoffs of disaggregating that line system in detail elsewhere, so I won't repeat the full pros-and-cons list here. If you're weighing vendor lock-in against operational complexity, our white paper on open line systems and disaggregation walks through five real deployment patterns, and Is hop-by-hop the future of optical transport networking? digs into what happens to ROADM path protection when Segment Routing and TI-LFA take over that job instead. For DCI purposes, the short version: a hub site with a dozen wavelengths terminating on a ROADM often keeps a traditional OLS, while a simpler point-to-point spoke between two facilities is a good candidate to run pluggable DWDM optics straight out of the router. Most DCI networks end up as a mix of both, not a wholesale replacement of one by the other.

Pluggable coherent optics: the piece doing the most work right now

400ZR and OpenZR+ made 400G DCI routine, and that history, from CFP2-ACO to CFP2-DCO to today's QSFP-DD modules, is covered well in our earlier piece on 400G-ZR and ZR+. What's changed since then is scale. The OIF finalized the 800ZR implementation agreement in late 2024, targeting single-span, amplified 80-120km DCI links, and Dell'Oro projects 800ZR/ZR+ could account for over a third of all IPoDWDM coherent pluggable revenue by 2026.

We covered the newest pluggable form factors, pluggable EDFAs, pluggable optical protection switching and POLS, in our 2026 optical networking trends recap, so I'll point there rather than re-explain them. The one addition worth flagging for a DCI-specific audience: security is moving into the pluggable too. Vendors including Marvell and Nokia now build MACsec encryption directly into their 800G coherent DSPs, encrypting traffic at line rate between facilities without a separate encryption appliance. For any organization moving regulated or sensitive data between data centers over leased or shared dark fiber, that's a meaningful architectural detail, not just a checkbox.

Why this becomes a budget conversation

DCI capacity has traditionally lived inside the network team's WAN or dark fiber budget, well below the level where a CFO or business unit leader ever looks. That is starting to change, for three reasons budget owners are noticing on their own.

GPU economics turn the network into a utilization problem finance already tracks. A distributed AI cluster is a large, sunk capital investment, and an interconnect that cannot keep pace leaves that investment sitting idle. That is a far easier case to make to a budget owner than "we need more bandwidth."

Power-driven site selection, covered above, means DCI cost now belongs in the real estate business case rather than a line item added after the lease is signed. A site picked for its available power, hundreds of kilometers from the nearest hub, carries an optical transport cost the original site economics may not have accounted for.

Converged architectures make the tradeoff quantifiable too: fewer chassis, fewer support contracts and less rack, power and cooling per site is a number that can sit next to a disaggregated OLS quote for direct comparison.

The forcing event is usually one of those three: a new AI cluster build, a capacity crunch discovered mid-planning or a new site whose lease never priced in transport. That is when networking, facilities and finance end up in the same meeting, and what they need from that conversation is not a technology preference. It's a comparison of total cost and risk across the options above.

Choosing the right architecture

There's no single right answer for interconnecting distributed data centers. The decision comes down to distance, wavelength count, how many sites need to drop traffic versus pass through it and how much of that complexity an operator wants to manage in the router versus a dedicated optical layer. What's changed is that 800ZR+ and its security and amplification options make it realistic to reconsider that decision, even for networks that settled on a traditional OLS years ago, and hyperscaler-driven scale is pulling that reconsideration forward faster than most roadmaps expected a year ago.

If you're evaluating where your DCI architecture stands today, WWT's converged optical and packet networking briefing and implementation services are good starting points for a conversation.