LPO, CPO and 800G DWDM: Where Each Fits in AI Networks
By Rena Erez | Technical Editor: Alex Surpin
TL;DR
AI is changing optical networking at more than one layer.
LPO (Linear Pluggable Optics) and CPO (Co-Packaged Optics) are designed to address power, density and signal-integrity constraints around high-performance switching. Coherent optics and 800G DWDM address a different requirement: transporting high-capacity traffic across longer fiber links, including campus, data center interconnect (DCI), metro and regional networks.
As networks move from 400G and 800G toward 1.6TbE, the focus is shifting from a single optical approach to choosing the right technology for each distance and function.
AI clusters are putting extraordinary pressure on network bandwidth, but raw capacity is only part of the problem. Power, switch density, fiber availability and the distance between compute locations are becoming just as important.
As AI clusters extend across sites, AI data center interconnect becomes part of the same capacity challenge, bringing coherent optics and 800G DWDM into the architecture.
That is why terms such as Linear Pluggable Optics (LPO) and Co-Packaged Optics (CPO) are appearing more often in discussions around AI network architecture.
Both are responses to a practical challenge: at very high speeds, moving electrical signals between the switch ASIC and the optical interface consumes increasing amounts of power and becomes harder to engineer.
LPO and CPO Address Power and Density at the Switch
With LPO, the optical module uses a simplified, more linear signal path with reduced or no retiming DSP compared with conventional pluggable optics. More of the signal conditioning is handled by the host system, potentially reducing module power, latency and complexity.
CPO takes integration further by placing optical engines in the same package, or in the immediate package vicinity, as the switching ASIC rather than using conventional front-panel pluggable modules.
These approaches are particularly relevant to high-density AI fabrics, where switch I/O power, thermal load, signal integrity and bandwidth density are major design constraints.
Neither LPO nor CPO removes the need for coherent transport where distance and optical-line requirements call for it.
As connections extend beyond the local AI fabric to campus, DCI, metro or regional links, the optical design problem changes.
When AI Traffic Extends Beyond the Local Fabric
AI infrastructure may span multiple halls, buildings, campuses and data centers. As those connections extend, architects need to consider the optical path as well as the switch interface.
They need to consider:
- available fiber and fiber type
- distance, loss and optical margin
- wavelength capacity and channel plan
- power and rack space
- service visibility and performance monitoring
- protection, restoration and fault isolation
- compatibility of existing DWDM infrastructure with 400G, 800G and future line rates
At these distances, coherent optics, 800G DWDM and dedicated optical transport become increasingly relevant.
The distinction is important.
LPO and CPO address how high-speed optical I/O is implemented around the switch. DWDM addresses how multiple high-capacity optical channels are transported over shared fiber infrastructure.
Where 800G DWDM Fits
For an operator building or expanding an AI data center interconnect, 800G can improve the economics and density of the optical transport layer.
An 800G wavelength can carry twice the payload capacity of a 400G wavelength, allowing operators to transport the same aggregate capacity with fewer coherent line interfaces. This can reduce optics count, rack space and power per transported bit. It does not automatically double total fiber capacity, which also depends on channel width, modulation, reach and the end-to-end optical design.
PacketLight's PL-8000G, for example, supports up to 10 x 800G transponders and 8 Tbps of capacity in 1U, with 100GbE, 400GbE and 800GbE client options. It also provides a dedicated demarcation point between the client equipment and the DWDM network.
That separation can be useful because the lifecycle of the optical transport network does not necessarily need to follow the lifecycle of the switching platform.
A data center or network operator may replace or upgrade switching platforms while retaining the underlying fiber and DWDM infrastructure. Likewise, an operator may introduce 800G wavelengths while continuing to carry existing 100G and 400G services.
The Architecture Is Becoming More Diverse, Not Less
There is unlikely to be one universal optical technology for every AI connection.
AI networks will use a mix of technologies according to reach, capacity and operating model:
retimed pluggable optics, LPO, CPO, coherent pluggables, dedicated 800G transponders and muxponders, and DWDM transport.
The right choice depends on where the connection sits in the network and what the operator needs from it.
Around high-capacity switching, power, thermal constraints, signal integrity, latency and bandwidth density are key considerations.
Across the optical path, fiber characteristics, channel plan, optical margin, amplification, protection, monitoring and operational control become increasingly important.
As 800G deployments accelerate and 1.6TbE interfaces emerge, network planning needs to cover the full path, not only the switch port.
As AI networks scale, no single optical technology will fit every part of the infrastructure. The practical task is to match each connection with the right balance of capacity, reach, power, fiber utilization and operational control.
Questions About LPO, CPO and AI Optical Networking
What is the difference between LPO and CPO?
LPO, or Linear Pluggable Optics, removes the retiming DSP used in conventional retimed pluggables and relies more heavily on the host SerDes and linear signal path. CPO, or Co-Packaged Optics, places the optical engines in or immediately adjacent to the switch package. Both approaches target lower power and higher bandwidth density, but they differ significantly in implementation, serviceability and system design.
Does CPO replace DWDM?
No. They address different architectural functions. CPO changes how optical I/O is integrated with the switching platform, while DWDM combines multiple optical wavelengths for transmission over shared fiber infrastructure. CPO and DWDM can therefore exist in the same end-to-end network.
Where does 800G DWDM fit in an AI network?
800G DWDM is particularly relevant when high-capacity traffic must move between data centers, campuses or other geographically separated compute locations. Compared with transporting the same aggregate client capacity over two 400G wavelengths, an 800G wavelength can reduce the number of coherent line interfaces. The total capacity available on the fiber still depends on the channel plan, spectral efficiency and optical design.
Are LPO and CPO alternatives to coherent optics?
Not directly. LPO and CPO describe how switch-facing optical I/O is implemented, while coherent optics describes a signaling and detection approach used for high-capacity optical transmission. They are not equivalent technology categories. In practice, LPO and CPO are most associated with power-efficient data center interconnects around the switching fabric, while coherent optics is widely used when reach and optical performance require it.
Why use a dedicated 800G transponder instead of putting coherent optics directly into a switch?
There is no single answer for every network. Router- or switch-hosted coherent optics can make sense for simple, qualified point-to-point Ethernet links. A dedicated transport platform can be more appropriate when the operator needs independent optical monitoring and fault isolation, service demarcation, protection, mixed client rates, an optical lifecycle independent of the switching platform, or OTN-based service functions where required.
Can existing DWDM infrastructure support 800G?
It depends on the existing optical infrastructure and the end-to-end optical design. Operators need to assess mux/demux passband and channel width, channel plan, ROADMs where present, amplification, fiber characteristics and available optical margin. In some networks, 800G can be introduced by adapting the optical layer or using a FlexGrid design rather than replacing the entire DWDM infrastructure.
What will 1.6T mean for optical network design?
As client interfaces move toward 1.6TbE, power, optical bandwidth and fiber planning become increasingly important. Operators will need to assess whether the underlying DWDM infrastructure, including mux/demux passbands and channel spacing, ROADMs where present, amplifier design, fiber characteristics and end-to-end optical margin, can support the required line-side signals and future capacity growth.