Data Center Interconnect (DCI): Solutions Overview

Last quarter a customer called about a 120 km data center interconnect (DCI) link between two of their facilities that kept dropping after roughly forty minutes of clean traffic. They had bought coherent modules rated for the distance, the dark fiber was their own, and the datasheet promised the reach. It still failed, repeatedly. When we pulled their span numbers, the link was sitting near 21 dB of insertion loss against a budget that left almost no headroom once connector aging and temperature drift were added back in. The modules tested fine on our bench. The plan was wrong.

 

That gap, between what a data center interconnect is supposed to do on paper and what it does on installed fiber, is where most of these projects actually struggle. Deciding to connect two sites is rarely the hard part. Picking the right optics for that connection, at the right reach, with enough margin to survive five years of cable-tray compression, almost always is. So this overview skips the textbook definition and walks through the solution classes the way an application engineer reasons about them: by distance, by loss budget, by what breaks in the field.

 

 

 

DCI reach does not scale smoothly. It steps. Each step changes the dominant constraint, and a solution that is obviously correct in one tier is wasteful or unworkable in the next. Here is how the tiers actually divide once you account for loss budget rather than marketing reach.

 

Distance tier	Typical span	Dominant constraint	Where the optics class lands
Campus / intra-site	under ~2 km	cost per port, density	grey QSFP-DD / OSFP, DAC, AEC
Metro point-to-point	~2–80 km	fiber loss, single span	grey + external DWDM, or 400ZR
Regional	~80–500 km	OSNR, amplification, ROADM passes	ZR+ / OpenZR+, often with line system
Long-haul / inter-regional	~500–2,000 km+	accumulated noise, dispersion	800ZR+ high-power coherent, transponder

 

The metro band is where the most expensive mistakes happen, because it is the band where a 400ZR-class coherent module looks like a drop-in answer. The OIF defined 400ZR for single-span links of around 80 km, targeting roughly 16 QAM at about 60 Gbaud inside a module power envelope near 15 W so it fits a QSFP-DD cage, per the OIF's published 400ZR implementation agreement. On a clean, short metro span a 400ZR DCI link is a genuinely elegant solution. The trouble starts when someone reads "80 to 120 km" as a guarantee. It is a ceiling under ideal loss conditions, not a promise on your fiber.

 

Push past the metro band and you are into regional territory, where ZR+ and OpenZR+ variants add output power, better forward error correction, and the multi-rate flexibility a routed optical network needs. At the far end, current top-of-range 800G coherent platforms reach on the order of 1,000 km at full 800 Gbps and roughly 2,000 km at reduced line rates, which is what makes hyperscale long-haul DCI over a single pluggable possible at all. Each tier earns its own solution; the failure pattern is forcing one tier's optics into the next tier's distance.

 

 

Underneath the distance tiers sit three solution classes, and most DCI architecture arguments are really arguments about which of the three to standardize on. The honest comparison is multi-dimensional, so it belongs in a table rather than buried in prose.

 

Solution class	Best reach fit	Cost & power posture	Operational reality
Grey optics + external DWDM / open line system	metro, short regional	low module cost, but you buy and run the line system	most moving parts; you own mux, amps, dispersion plan
Coherent pluggable (IP-over-DWDM, 400ZR / 800ZR)	metro to long-haul, point-to-point	lowest space and power per bit; optics live in the router	simplest topology, hardest interoperability story
Transponder / thin transponder	any, esp. multi-service or ROADM	highest capex, decoupled from the router	operational consistency across a mixed optical network

 

That first row also settles the DCI DWDM versus dark fiber question that comes up on almost every metro build: if the dark fiber is yours, you light it with your own DWDM or open line system and keep the wavelength plan in-house; if you are leasing lit wavelengths instead, that line system belongs to the carrier and your optics have to live inside whatever it provides.

 

The live industry debate is between coherent pluggables and transponders. Hyperscalers pioneered the IP-over-DWDM model, dropping 400ZR and now 800ZR coherent optics straight into router faceplates to strip out the space, power, and cost of a separate transport shelf, and large operators continue down that path for their point-to-point coherent data center interconnect links (Data Center Knowledge). The clean takeaway is that pluggable coherent wins on cost and footprint. That conclusion holds, with a condition most vendors skip over: it holds cleanly for single-span, point-to-point links you control end to end. The moment a communications service provider needs to carry that wavelength across a multi-vendor line system, through ROADM nodes, alongside legacy services, the operational consistency of a thin transponder often beats the raw economics of stuffing the optic into the router. We get into the muxponder-versus-transponder mechanics that drive this in our muxponder and transponder breakdown; for now the point is that the "obvious" pluggable answer reverses in exactly the environments where most enterprises actually operate.

 

 

Here is the variable that decides whether your DCI optical transceiver selection survives contact with real fiber: link loss budget, not the reach printed on the box. A coherent interface tolerates a finite amount of optical loss and noise before its DSP can no longer recover the signal cleanly, and that tolerance is what the standard actually specifies. The OIF wrote 400ZR against two distinct link types, an amplified noise-limited link and an unamplified loss-limited link, precisely because the same module behaves differently depending on which regime you are in (OpenZR+ MSA). Translate that into field terms: an unamplified 75 km span with clean splices and good connectors can be well inside budget, while a "shorter" 60 km span with a dozen patch panels, an aging ODF, and two extra mated pairs can sit out of budget and fail intermittently.

 

 

So the planning number that matters for a metro or regional DCI link is your measured span loss with margin, not the kilometer figure. Our standard recommendation before any coherent deployment is an OTDR and loss characterization on every segment, not a sample, plus a margin allowance for connector aging and temperature drift. The expensive version of skipping this is the 120 km link from the opening of this page. The cheap version is two hours with a power meter. One more place where the gap between the catalogued reach and the deployable reach quietly decides the project.

 

 

"400ZR" means something specific. "ZR+" frequently does not. The base 400ZR interface is a single, interoperable OIF implementation agreement, and two compliant modules from different makers will talk to each other across the specified span. "ZR+" grew up afterward as a catch-all for higher-power, longer-reach, multi-rate coherent pluggables, and the term now covers everything from genuinely interoperable OpenZR+ implementations to vendor-specific extensions that only interoperate with their own line systems. Several distinct technical definitions travel under the same two-character suffix, which is how procurement teams end up with two "ZR+" modules that refuse to link.

 

Treat the label as a question, not an answer. Before you commit to a cross-vendor regional data center interconnect built on ZR+, confirm three things: the exact implementation agreement each module claims, whether both ends share the same FEC and modulation profile, and whether the line system's amplification and filtering plan matches what the module expects. If those line up, ZR+ delivers the reach extension it promises. If they do not, you have bought two incompatible optics that each work perfectly in isolation, which is the most frustrating failure mode in the entire interconnect catalog because nothing is technically broken.

 

 

Coherent optics run hot, and a metro DCI module that draws a few extra watts is not a rounding error at the rack level. The reason 400ZR was constrained to roughly a 15 W envelope was to make it survivable inside a QSFP-DD cage in a densely loaded router; that thermal ceiling shaped the whole standard, a constraint the OIF was explicit about when it set the power target. Stack high-power coherent modules into a faceplate that was provisioned for grey client optics and you can push the cage past its cooling design, which is exactly why router vendors warn that third-party high-power coherent and ZR+ modules can cause thermal damage to host equipment, with liability landing on the operator who installed them.

On our side, every coherent and high-speed module design goes through sustained thermal-cycling qualification at 85 °C for 2,000 hours, with laser bias-current drift tracked as the primary aging indicator and units pulled from the line above a fixed drift threshold. The number that should appear in your selection conversation is the module's real power class against the host's per-cage thermal limit, because at 800G densities an 800ZR or ZR+ data center interconnect packs the most reach into the fewest watts, and the difference between a 16 W and a 22 W coherent module decides whether a rack stays in its envelope or trips a thermal alarm at two in the morning.

 

 

Different optics, same handful of root causes. Pulling from our own RMA and field-support records, here are the patterns that account for most "the link won't come up" tickets on inter-site deployments, and what actually fixes each.

 

The first is over-distance passive copper at the short end of a campus interconnect. A customer once ordered 800G passive DAC assemblies in 3-meter lengths based on an old rack layout; link training failed on over 60% of those ports, not because the cable was defective but because the copper loss at 112G per lane simply will not carry that far. Active electrical cable for the 3-to-7-meter gap and optical modules beyond it stabilized the deployment within a week.

 

The second is contamination, and it is the quiet majority. Somewhere in the range of two-thirds of high-speed link failures we investigate trace back to connector and end-face contamination rather than transceiver faults; a recent batch of forty 400G modules came back as alleged failures and every one passed our regression clean, with the real culprit being particulate on the trunk-cable end-faces. A fiber microscope and a proper cleaning kit prevents more outages than any module upgrade.

 

The third and fourth are the planning errors this page has already named: sizing a coherent link by nameplate kilometers instead of measured loss budget, and dropping a baseline 400ZR optic into a multi-hop ROADM network it was never specified for. Both look fine in a lab point-to-point test and fail once they meet the real optical path. If you want the structured way we walk a span before committing to optics, it is laid out in our optical network planning guide.

 

 

Your situation	Sensible default	Watch for
Two sites, under 80 km, your own dark fiber, point-to-point	400ZR coherent pluggable, IP-over-DWDM	measured span loss vs budget
80–500 km, point-to-point, you control both ends	ZR+ / OpenZR+ coherent pluggable, possibly amplified	which "ZR+" definition; host thermal class
Regional or long-haul across a multi-vendor / ROADM line system	transponder or thin transponder	operational consistency over raw module cost
Metro, mixed legacy services, no in-house optical team	grey optics + managed external DWDM	total line-system ownership cost
500–2,000 km, single-operator backbone	800ZR+ high-power coherent	accumulated OSNR, power and cooling

 

The trap is reading a row as the verdict rather than the starting point. Take the second row: ZR+ is the obvious default for an 80–500 km point-to-point link, right up until one end terminates behind a leased ROADM you don't control, and the right answer quietly flips to a transponder even though every distance and cost number still says pluggable. We have watched a "settled" ZR+ plan unravel in exactly that handoff. Which exception is hiding in your span, your switch platform, or your operations model is the one thing this table cannot see from here.

 

 

The throughline of every section here is the same: the data center interconnect decision is an optics-and-fiber decision long before it is a bandwidth decision, and the projects that go smoothly are the ones that characterized the span and pinned down the power and interoperability story before ordering modules. For context on the kit behind that promise, we ship coherent and grey optics from 1G SFP up to 800G OSFP, with 1.6T OSFP224 in qualification, and every coherent module is validated across more than a dozen switch and router platforms before it leaves the line. If you are mapping a specific link, the most useful thing we can do is check your distance, measured loss, and switch or router platform against the right reach class, then send real specs and samples to qualify before you scale. You can start from our DCI and OTN transport product line for the coherent and transponder side, or the DWDM and CWDM range if you are building the wavelength plan around grey optics. Either way, the conversation we have every week is the same one this page is trying to start: not "how far does it go," but "how far does it go on your fiber."