Coherent DWDM in the 800G Era

An artificial intelligence (AI) training job split across two buildings 100 km apart hits a latency wall long before it runs out of bandwidth. That constraint is reshaping how data centers get interconnected, and it is why AI is now the largest single driver of demand in optical networking.

Coherent dense wavelength division multiplexing (DWDM) has kept pace. 800G ZR and ZR+ modules ship in volume from six or more vendors, 1.6T silicon is being announced ahead of the standards that will govern it, and the convergence of routing and DWDM is now simply how large networks get built. Here is where each generation stands, what it costs to run and what to spec.

How the optic ended up in the router

Coherent optics pair advanced optics with a digital signal processor (DSP) to send and receive complex light wave modulations. The shift that mattered was packaging: once the DSP shrank enough to fit inside the pluggable module itself, DWDM transport could run straight out of a router and the standalone transponder shelf became optional. 400G-ZR and ZR+ in QSFP-DD put a tunable coherent wavelength in the same faceplate as a grey 100G optic, and converging DWDM with routing stopped being an engineering argument and became a purchasing decision.

Figure 1: The end of the transponder. A traditional optical layer of transport shelves and transponders collapses into pluggable coherent optics seated directly in the router. ZR covers cost-effective point-to-point spans; ZR+ covers multi-span regional reach. 1.6T will follow the same model.

The business case, in numbers

The economics are simple enough to work out on a napkin, and we have. Delivering 400G across the wide area the traditional way means a transponder line card in a dedicated optical chassis — roughly 330 watts. The converged equivalent, a DWDM-capable pluggable in the router, draws about 75 to 85 watts. That is a 77 percent power reduction in a single function, and more than 4,000 kilowatt-hours a year per 400G link. Published modeling of routed optical architectures puts the broader savings at up to 45 percent on power and up to 70 percent on floor space.

The operational savings compound on top. One control plane. One management system. One team running both IP and optical transport instead of two.

The market has already priced this in. Cignal AI puts standalone coherent pluggable revenue — modules sold directly by suppliers to network operators — at $1.8 billion in 2025, up 37 percent, and forecasts more than $5 billion by 2029 at a 30 percent compound annual rate. Pluggables already carry more than half of all WDM bandwidth shipped. A Heavy Reading survey in October 2025 found 58 percent of service providers planning 800G pluggable deployments between 2026 and 2028, which is usually the figure that lands hardest in an internal business case.

400ZR, OpenZR+ and the alphabet soup

The naming trips people up. The Optical Internetworking Forum (OIF) created 400ZR for short-reach data center interconnect (DCI) inside a 15 watt budget. The Open ROADM Multi-Source Agreement (MSA) specified a longer-reach variant with open forward error correction (oFEC) and selectable line rates, which became ZR+. The two were combined into OpenZR+. Hold onto that structure — a tight OIF short-reach standard plus a longer-reach ZR+ profile — because it repeats at 800G and again at 1.6T.

800G ZR and ZR+: already here

800G is not a roadmap item. The OIF 800ZR implementation agreement landed in October 2024, defining short-reach 800G as an 80 to 120 km amplified span. The OpenZR+ MSA 800G white paper followed in July 2025, extending 800G to regional and long-haul applications with selectable rates and probabilistic shaping. Modules are generally available across the major optical vendors, and the ecosystem went from early access to mainstream in under two years.

Three things changed from 400G. Baud rate roughly doubled, to about 118 Gbaud from roughly 60 Gbaud, and the DSP moved to 3 or 5 nanometer silicon depending on the vendor, which is what keeps a coherent module inside the thermal budget of a router faceplate.

Probabilistic constellation shaping (PCS) is the second. By biasing the signal toward lower-amplitude symbols, which resist noise better, it turns the capacity-versus-reach trade-off into a continuous dial rather than discrete jumps between modulation formats. It is much of why the 800G reach figures below look as good as they do.

Third, and easiest to overlook: 800ZR uses an 800GAUI-8 host interface, eight lanes of roughly 112G PAM4, which is already standard on shipping 800GbE routers and switches. Convenient timing, and a large part of why 800G coherent drops into existing platforms so cleanly.

One practical note for the requirements document: unlike 400G, which settled on QSFP-DD, 800G ships in both QSFP-DD800 and OSFP. The larger OSFP body buys thermal headroom, which matters for the higher-power ZR+ configurations.

What the standards define, and what they do not

Every coherent module trades capacity against distance — the higher the line rate, the shorter the run. The published standards define the rates, the modulation and the forward error correction precisely. What they mostly do not define is distance.

Sources: OIF 400ZR and 800ZR implementation agreements; OpenZR+ MSA 800G white paper (July 2025). 1.6T is omitted because no implementation agreement has been published.

AI: the new demand engine

Large-scale model training runs across thousands of graphics processing units at once. For years that traffic stayed inside one building. Then clusters outgrew what any single building could supply — power density, cooling capacity and floor space pushed operators to distribute training across multiple sites, sometimes hundreds of kilometers apart. Every link between them has to carry terabits with minimal latency and complexity. 800G ZR and ZR+ answer that directly, and the internet protocol over DWDM (IPoDWDM) model fits how hyperscalers prefer to operate: merchant silicon, software-defined, no optical specialist required. Hyperscalers have led adoption.

Reach is now a function of the workload, not the map

This is the part that changes how the specification gets written. Distance used to be a geography question — how far apart are the two sites? For AI fabrics it is a function of what the workload tolerates, and the three main patterns tolerate wildly different amounts. Synchronous training is the tightest by a wide margin, because every step waits on the slowest link. Inference is looser. Asynchronous replication between regions is looser still. No standards body has put numbers on these thresholds, and the figures circulating from conference panels are contested, so treat any specific kilometer count you are quoted as a starting hypothesis rather than a limit.

The consequence is concrete regardless of the exact numbers. Long regional reach is worth nothing to a synchronous training fabric that breaks well before it, while asynchronous replication can use every kilometer available. The optical decision now depends on a compute architecture decision, so it has to be made alongside the platform team rather than downstream of them. That is a change in who is in the room, not just a change in the specification.

What is next: 1.6T

1.6T coherent is close but not here. 1600ZR and 1600ZR+ are both active OIF projects, but neither implementation agreement has been published as of this writing — check the OIF implementation agreements page before relying on any 1.6T specification. Multiple vendors announced 1.6T coherent pluggables during 2026, which means silicon is running ahead of the standard and early modules are validated against draft baselines. Treat any claim that 1.6T shipped in 2025 carefully: the 1.6T optics that reached production that year were client-side direct-detect datacom transceivers, a different product class from coherent DWDM. The form factor is not settled either.

The convergence is no longer a forecast

Hyperscale cloud providers moved first, retiring standalone DWDM transponder shelves in favor of coherent pluggables directly in router ports. Service providers are following, because the economics above are increasingly difficult to argue against. Combined with segment routing, the converged IP and optical layer is becoming standard for DCI, metro aggregation, 5G transport and now distributed AI compute.

The question is no longer whether to converge IP and DWDM. It is how fast to move, and what to write into the specification when you do. Three things worth settling before the next procurement cycle:

• Is 400G still right for this link? For anything going live in 2027 or later, 800G is increasingly the default.
• Do you know the workload's latency ceiling? Specify reach after that conversation, not before it.
• What is your per-slot power and thermal budget? At 40 watts a module, this constrains platform choice in ways 400G never did.

To see these architectures side by side, reach out or take a look at our converged optical and packet networking briefing.