Optical DCI vs Ethernet DCI

Data center interconnect (DCI) is the technology that connects two or more data centers across campus, metro, or regional distances, enabling them to operate as a unified resource pool. As AI workloads scale beyond the power and space available in any single facility, and as cloud platforms distribute across multiple availability zones, DCI has shifted from a background infrastructure concern to a primary determinant of system performance and cost.

Two architectural approaches dominate DCI design: optical DCI and Ethernet DCI. The distinction is not about whether fiber is used—both use fiber. It is about where the intelligence resides and how the signal is transported. Optical DCI relies on coherent optics, dense wavelength division multiplexing (DWDM), and optical transport platforms to carry data across long distances with minimal electrical processing. Ethernet DCI relies on Ethernet switches and routers, using PAM4 direct-detect optics for shorter reaches and coherent pluggables for longer ones, with packet processing at every node.

The choice between them is not binary. Most real DCI networks combine both approaches, assigning each layer of the network to the technology that handles it most efficiently. But understanding the trade-offs—in reach, capacity, cost, power, and operational complexity—is essential for designing DCI that can scale with AI infrastructure demands.

1. What DCI Must Accomplish

Before comparing architectures, it helps to establish what a DCI network is actually required to do. The requirements are more demanding than they might appear.

• High throughput: DCI carries both north-south traffic between sites and east-west traffic for distributed systems. AI training and inference generate massive east-west flows that must traverse DCI links without becoming a bottleneck.

• Low and stable latency: Even small increases in latency can affect application performance, especially in storage replication, cluster orchestration, and distributed AI training.

• Scalability: Traffic growth comes from more servers, higher utilization, and new workloads. DCI must scale without forcing disruptive redesigns.

• High availability: Outages or degraded performance directly impact customer-facing services. DCI networks must provide carrier-grade reliability with fast protection switching.

• Efficient operations: Network teams need predictable provisioning, monitoring, and troubleshooting workflows across what may be hundreds of wavelengths and thousands of links.

Optical networks map well to these requirements because they provide high capacity, long reach, and predictable performance while enabling scalable expansion through standardized optical interfaces and transport methods. Ethernet networks map well to DCI because they leverage the same switching technology, operational tools, and staff expertise that already exist inside the data center.

2. The Architectural Divide: Optical Layer vs Packet Layer

The fundamental difference between optical DCI and Ethernet DCI is the layer at which the network operates.

2.1 Optical DCI: The Transport Layer

Optical DCI builds the interconnect from the optical layer up. Data center traffic, typically Ethernet-framed, is mapped into optical wavelengths by transponders or coherent pluggable modules. Those wavelengths are multiplexed onto fiber using DWDM, amplified by erbium-doped fiber amplifiers (EDFAs) at intervals, and routed through optical add/drop multiplexers or optical cross-connects. The optical layer is responsible for transport—moving bits from one site to another—while the client layer (switches and routers) handles packet forwarding.

This separation of concerns allows the optical layer to be optimized independently for reach, capacity, and reliability. Optical transport platforms provide functions that Ethernet switches do not: wavelength management, optical power monitoring, amplifier control, and protection switching at the optical layer. An optical network can reroute traffic around a fiber cut in milliseconds without involving the packet layer at all.

2.2 Ethernet DCI: The Packet Layer

Ethernet DCI builds the interconnect from the packet layer. Data center switches and routers are connected directly using Ethernet optics—PAM4 direct-detect for short reaches within a campus, or coherent pluggables for longer metro and regional links. The switch or router performs packet processing, routing, and congestion management at every hop.

When coherent pluggable optics are inserted directly into router or switch ports, the IP device becomes part of the optical transport layer, handling wavelength tuning, forward error correction, optical monitoring, and coherent transmission itself. This is called IPoDWDM (IP over DWDM), and it reduces hardware layers, saves rack space, and simplifies cabling by eliminating the separate transponder shelf. The operational boundary, however, collapses: the routing team now owns wavelength, OSNR, and FEC behavior in addition to packet forwarding.

3. Coherent Optics: The Enabling Technology for Long-Reach DCI

Coherent optics is what makes optical DCI possible beyond campus distances. Unlike PAM4 direct-detect, which measures only the intensity of light, coherent detection recovers the full optical field—amplitude, phase, and polarization—by mixing the received signal with a local oscillator laser. This allows coherent systems to use higher-order modulation formats such as 16-QAM, which encode four bits per symbol, and dual-polarization transmission, which doubles capacity per wavelength.

The practical consequence is spectral efficiency. A coherent 400G ZR link carries 400 Gbps over a single wavelength on a single fiber pair. A PAM4 400G link requires eight parallel fibers, each carrying a single lane. In fiber-constrained environments, coherent's fourfold advantage in spectral efficiency is decisive.

Coherent DSPs also compensate for chromatic dispersion, polarization mode dispersion, and carrier phase noise entirely in the digital domain. This eliminates the need for dispersion-compensating fiber or optical compensation modules, and it allows coherent links to span 80 to 120 kilometers without regeneration in standard DCI configurations, and hundreds of kilometers with appropriate amplification. PAM4 direct-detect has no mechanism to compensate for dispersion at the receiver; its reach shrinks as data rates increase, falling to under 10 kilometers at 200G per lane.

4. IPoDWDM: Bridging the Two Architectures

The traditional optical DCI architecture uses separate platforms for the packet layer and the optical layer. A router or switch sends a short-reach client signal to an external transponder or muxponder, which converts the signal into a wavelength suitable for DWDM transmission. The optical transport platform then multiplexes, amplifies, and routes that wavelength across the fiber plant.

IPoDWDM changes this model by inserting coherent pluggable optics directly into router or switch ports. The IP device becomes the optical transport endpoint, handling wavelength tuning, FEC, optical monitoring, and coherent transmission itself. This eliminates the transponder shelf, reducing hardware layers, rack space, and cabling. It also speeds service turn-up, because a new wavelength can be provisioned by inserting a pluggable module rather than installing and configuring a transponder card.

The trade-off is operational. With a transponder shelf, the IP team and the optical transport team can work independently, each with its own fault domain and management plane. With IPoDWDM, those two worlds collapse into one device. The routing team now owns wavelength planning, OSNR management, and FEC behavior—responsibilities that traditionally belonged to the optical team.

IPoDWDM is not automatically the best choice for every link. It requires compatible host platforms with sufficient power and cooling, proper coherent module support, and careful optical design. For simple point-to-point links over dark fiber, 400ZR or 800ZR pluggables provide a clean and cost-effective solution. For complex multi-service transport across many sites, a transponder or muxponder platform often remains the safer and more flexible design.

5. Open Line Systems and Disaggregation

The optical DCI landscape has evolved toward disaggregation. Traditional optical transport platforms were monolithic: a single vendor supplied the transponders, amplifiers, multiplexers, and management system as an integrated package. Open Line Systems (OLS) separate the optical line system—the amplifiers, multiplexers, and monitoring components—from the coherent endpoints.

This separation allows network architects to select transceivers and line system components independently, and to upgrade each layer on its own schedule. An Open Line System introduces programmability and centralized wavelength management, which is particularly valuable for multi-node or mesh DCI topologies. Standardized interfaces and optical-layer disaggregation give DCI network architects more flexibility when selecting and upgrading network components.

Two main approaches to OLS have emerged. Modular optical platforms organize transport, line-system, service, and monitoring capabilities into configurable modules or service cards, allowing incremental expansion and support for complex topologies. Integrated OLS platforms consolidate core optical line-system components—muxponders, DWDM multiplexers, EDFAs, and optical supervisory channel modules—into a compact chassis, prioritizing deployment simplicity and operational efficiency over maximum flexibility.

6. Ethernet DCI in Practice

Ethernet DCI is not a single architecture. It spans a range of technologies depending on distance and capacity requirements.

6.1 Campus and Short-Reach DCI

For links within a campus or between adjacent buildings—typically under 2 kilometers—Ethernet DCI uses standard PAM4 direct-detect optics. A 400G or 800G Ethernet switch port connects to a module that transmits PAM4 signals over multimode or single-mode fiber. The DSP performs equalization and forward error correction, and the packet layer handles routing and congestion management. This approach leverages the same switching technology, operational tools, and staff expertise that already exist inside the data center.

6.2 Metro and Regional DCI

For links spanning metro or regional distances—2 to 80 kilometers—Ethernet DCI increasingly uses coherent pluggable optics in router or switch ports. A 400G ZR or 800G ZR module converts the Ethernet signal into a tunable DWDM wavelength, which is then multiplexed onto fiber and amplified as needed. This is IPoDWDM in practice: the router or switch becomes the optical transport endpoint, and the separate transponder shelf is eliminated.

The operational simplicity of this approach is compelling. A single device handles both packet forwarding and optical transport, reducing rack space, power consumption, and management points. But it also means that the routing team must understand optical parameters—OSNR, chromatic dispersion, launch power—that were previously the domain of optical engineers.

7. Reach, Capacity, and Application Mapping

The choice between optical DCI and Ethernet DCI depends primarily on distance and capacity requirements. The following table maps common DCI scenarios to recommended approaches.

The pattern is clear. Ethernet DCI dominates the shortest reaches, where PAM4 direct-detect provides adequate performance at the lowest cost and power. As distance increases, coherent pluggables extend Ethernet DCI into the metro range. Beyond 80 kilometers, optical DCI with transponders and Open Line Systems becomes the standard approach, because the complexity of long-reach transmission is better managed in a dedicated optical layer than in a router port.

8. Cost and Power Considerations

The cost comparison between optical DCI and Ethernet DCI is more nuanced than module price alone.

8.1 Capital Expenditure

For short campus links, Ethernet DCI using PAM4 optics is less expensive. A 400G PAM4 module costs hundreds of dollars, while a 400G ZR coherent module costs several thousand dollars. The Ethernet switch or router already exists; the DCI function is added by inserting a module.

For metro and regional links, the comparison shifts. Ethernet DCI with IPoDWDM eliminates the transponder shelf, reducing capital expenditure. Optical DCI with a separate transponder and Open Line System requires additional hardware—transponders, amplifiers, multiplexers, and management systems—but provides reach and margin that pluggable optics cannot match over longer distances.

Dark fiber presents a compelling economic case for optical DCI at scale. Above 400G, dark fiber with self-deployed optical transport presents a three-year total cost of ownership saving of 48 percent versus Carrier Ethernet and 55 percent versus wavelength services. The availability of dark fiber is leading a growing number of operators to deploy their own DCI networks, giving them control over their network from end to end.

8.2 Power Consumption

Optical DCI reduces power consumption by eliminating electrical processing at intermediate nodes. In an optical transport network, wavelengths pass through optical amplifiers and ROADMs without optical-electrical-optical conversion. In an Ethernet DCI network, every switch or router in the path performs OEO conversion and packet processing, consuming power at each hop.

Coherent pluggable modules consume more power than PAM4 direct-detect modules—approximately 1.3 to 1.5 times at equivalent data rates—but they eliminate the power consumption of transponder shelves and intermediate packet processing. At the system level, the power comparison depends on the number of hops and the reach required.

An IP+optical fusion approach for intelligent computing interconnect has demonstrated a 40 percent reduction in cost per bit, a 35 percent reduction in power consumption, and a 20 percent reduction in node latency compared to traditional "router + transport equipment" DCI solutions.

9. Where Each Approach Excels

9.1 Optical DCI Strengths

• Long reach: Coherent transmission with amplification spans hundreds of kilometers without regeneration.

• Spectral efficiency: DWDM and coherent modulation carry multiple terabits per fiber pair.

• Protocol transparency: The optical layer is agnostic to Ethernet, InfiniBand, or other client protocols.

• Independent fault domains: The optical layer can be managed and protected independently of the packet layer.

• Deterministic latency: No packet buffering or processing at intermediate nodes.

9.2 Ethernet DCI Strengths

• Operational familiarity: Leverages existing switching technology, tools, and staff expertise.

• Simplified architecture: IPoDWDM eliminates the transponder shelf, reducing hardware and management points.

• Lower cost for short reaches: PAM4 optics are significantly less expensive than coherent for campus distances.

• Faster service turn-up: New wavelengths can be provisioned by inserting a pluggable module.

• Integrated packet and optical management: Fewer management systems to integrate.

10. The Convergence Trajectory

The boundary between optical DCI and Ethernet DCI is blurring. Coherent pluggable optics are bringing optical transport functions into Ethernet switches and routers, while open line systems are disaggregating the optical layer into standardized components that can be managed alongside packet infrastructure.

Several trends are shaping this convergence:

• Higher-speed coherent pluggables: 800ZR and 1.6T coherent modules are entering the market, extending IPoDWDM into longer reaches and higher capacities.

• Coherent-lite for campus: Simplified coherent DSPs optimized for 2–20 kilometer campus links are bridging the gap between PAM4 and full coherent.

• Open optical networks: Standardized interfaces and optical-layer disaggregation allow components from different vendors to interoperate, reducing vendor lock-in.

• AI-driven DCI demand: AI training and inference require DCI that can support massive east-west traffic with low latency and high reliability, driving adoption of both optical and Ethernet DCI technologies.

The future of DCI is not a choice between optical and Ethernet. It is a layered architecture that uses Ethernet DCI for campus and metro connectivity, optical DCI for regional and long-haul transport, and IPoDWDM to bridge the two. Understanding where each approach performs best—and how they can be combined—is the key to building DCI that scales with AI infrastructure demands.

11.Conclusion

Optical DCI and Ethernet DCI represent two architectural approaches to connecting data centers. Optical DCI builds the interconnect from the optical layer, using coherent optics, DWDM, and optical transport platforms to achieve long reach, high spectral efficiency, and protocol transparency. Ethernet DCI builds the interconnect from the packet layer, using Ethernet switches and routers with PAM4 direct-detect for short reaches and coherent pluggables for longer ones.

The choice between them depends primarily on distance and capacity requirements. Ethernet DCI dominates campus and metro reaches, where PAM4 and coherent pluggables provide adequate performance at lower cost and with greater operational simplicity. Optical DCI dominates regional and long-haul reaches, where the complexity of long-reach transmission is better managed in a dedicated optical layer than in a router port.

IPoDWDM has emerged as a bridge between the two, inserting coherent pluggable optics directly into router and switch ports to eliminate the transponder shelf and collapse the IP and optical layers into a single device. Open Line Systems have disaggregated the optical layer into standardized components, giving network architects more flexibility in selecting and upgrading network components.

As AI workloads drive demand for higher bandwidth, lower latency, and greater scalability, both approaches will continue to evolve. The organizations that understand the trade-offs—and design their DCI architectures to match the requirements of each link—will be best positioned to build networks that support the next generation of AI infrastructure.

12.Q&A