Introduction
In Cignal AI’s discussions with operators and vendors over the last month, it became clear that CPO is moving closer to real deployment, largely due to the explosion in AI bandwidth requirements. As we look over the horizon at the next generation (or two) of AI buildouts, CPO looks inevitable–at least for some parts of the network.
However, CPO is not yet imminent. Technical challenges must still be overcome, viable alternative solutions are available in most applications, and business and operational issues need to be addressed. Small amounts of CPO may start to appear in 2026, but real deployment at scale looks more likely to arrive in 2027/8 or later.
This report dives deeper into CPO for insight on the technology and applications, the benefits and issues, its impact on pluggable optics, and Cignal AI’s predictions for CPO’s future.
Download PDF VersionKey Takeaways
- Large-scale CPO deployment is still 3-5 years away, although initial commercial trials may commence in 2026. The technology required to support CPO already exists, but the ability to manufacture at high volume will take until at least 2027. Acceptance by hyperscalers in the scale-out network will take even longer than that and may be delayed until 400G/lane makes the use case clearer.
- CPO has undeniable benefits, but there are also still a number of open issues that could keep it out of mass production for years, especially in parts of the network where current solutions are working well, such as scale-out.
- Initial CPO deployments will be proprietary, allowing for innovation before standardization. This leaves the door open to both established players and startups to apply their technologies into different applications with different customers.
- CPO’s main applications are within the scale-up and scale-out AI networks. Scale-up is where there is pain (bandwidth, density, power), but that network is currently served by copper, which beats optics in power and cost and will therefore be hard to dislodge. Scale-out is already 100% optical and looks like a good place to initiate CPO deployments, but there are existing solutions working well (pluggables) and improved solutions that can be adopted (LPO/LRO). As a result, there is less immediate need for CPO.
- Initial deployments will result from a large operator or system builder (hyperscaler or Nvidia) utilizing CPO for a specific application and will start in the scale-up network as an alternative to copper interconnect before migrating to the scale-out network.
- CPO’s impact on pluggable optics will be nonexistent at first, as initial applications will replace copper and not fiber. Long-term (more than 5 years), CPO will begin to cannibalize pluggables in scale-out applications, but there are enough connectivity requirements to keep pluggable vendors in business even then, if perhaps only at a lower scale.
Table of Contents
Clients log in to access full report