Packaged optics, known as Co-Packaged Optics or CPO, brings optical engines closer to the network processor to reduce the electrical distance between the fiber and the chip. Corning argues that this change will be necessary to sustain the growth of AI clusters, but warns that replacing pluggable transceivers alone is not enough: connectors, special fibers, trays, and optical arrays must be designed as a single integrated system.
The essentials of co-packaged optics in 20 seconds
- CPO moves the optical conversion closer to the network ASIC or GPU.
- Reduces electrical traces on copper inside the equipment.
- Can improve energy efficiency, bandwidth density, and latency.
- A system can integrate over 1,000 fibers with extremely precise alignments.
- Deployment requires manufacturing and validating the entire internal optical pathway as a single unit.
In a typical data center architecture, fiber reaches a transceiver installed at the switch front panel. The module converts the optical signal to electrical, and data travel over the board’s copper traces to reach the application-specific integrated circuit, or ASIC.
This route can be about 30 centimeters. Though small, it becomes problematic as channel speeds and total switch capacity increase. Signal degradation, interference, and the need for more digital processing to compensate for losses become significant issues.
Co-packaged optics change where the conversion occurs. Instead of placing the transceiver on the front panel, small optical engines are positioned near the switching ASIC, inside the same package or very close to it. The fiber enters deeper into the equipment, and the electrical segment is shortened to just a few millimeters.
Why copper starts to limit AI switches
Clusters used for training and running AI models need to move large volumes of data between GPUs, accelerators, memory, and storage.
GPU performance becomes irrelevant if it spends too much time waiting for data from other nodes. Therefore, internal networks are evolving from 25.6 Tb/s and 51.2 Tb/s switches toward 102.4 Tb/s platforms.
However, increasing switch capacity also requires boosting the speed of each electrical connection. In traditional designs, signals pass through connectors, printed circuit traces, and other components before reaching the optical module.
To keep signals usable, digital signal processing circuits, known as DSPs, are used. These correct errors and compensate for channel losses but add power consumption, heat, and cost.
Broadcom notes that this problem worsens with SerDes interfaces at 212 Gb/s using PAM4 modulation. The manufacturer claims that placing optics near the ASIC reduces path losses and can save over 3.5 times in energy compared to certain configurations with pluggable optics and DSP. They also estimate optical cost per bit reductions and bandwidth densities exceeding 1 Tb/s per millimeter, although these figures depend on the specific platform.
CPO does not eliminate optical-electrical conversion; it relocates it. The optical engine continues to receive and transmit signals but does so almost directly beside the processor that uses them.
An intermediate architecture called Near-Packaged Optics (NPO) exists. In this setup, the optical engines are very close to the ASIC but do not share the same package. This separation can simplify manufacturing and maintenance but results in a slightly longer electrical path.
Over 1,000 fibers inside a single device
Bringing optics further inside the switch solves part of the electrical issue but introduces new mechanical and optical challenges.
A CPO system may need to organize more than 1,000 fibers within a very limited space. Each fiber must reach a photonic integrated circuit, known as a PIC, with alignment tolerances less than a micron—one thousandth of a millimeter, according to Corning.
Fibers cannot be placed arbitrarily. Excessive bending increases losses, and poor organization complicates manufacturing and can put stress on connections during the equipment’s lifetime.
Design must balance four key requirements:
- Minimized insertion losses;
- High fiber density;
- Mechanical and thermal durability;
- Repeatable assembly at high volumes.
These goals often conflict. A more compact connector increases density but complicates assembly. Rigid fibers can improve mechanical performance but require more space to bend. Reusable, detachable connections ease installation but add optical interfaces that could introduce losses.
Therefore, Corning advocates for integrated design of front connectors, special fibers, internal trays, and fiber arrays.
Fiber Array Units (FAUs) group multiple fibers and align their cores with the photonic waveguides on the chip. Precision is critical: even minimal deviations can reduce light coupling into or out of the circuit.
Corning has developed bend-resistant fibers for short internal runs, along with pre-assembled optical trays and FAU sets ready to be installed adjacent to the chip. These components are bundled into their Optical Management Solution—a commercial package that customizes the optical pathway to each client’s chassis and ships as a tested, integrated unit.
Integration improves performance but impacts repairability
Pluggable transceivers offer a clear advantage: they can be removed and replaced without replacing the entire switch.
In a CPO architecture, much of the optics are integrated within the system and are not designed for on-site repair by data center technicians. An internal failure might require replacing a larger assembly or even the entire unit.
This has been a primary argument against early adoption of CPO. Operators are accustomed to swapping individual optical modules and having compatible spare parts from multiple vendors.
Proponents of the integrated approach argue that it reduces connectors, electrical paths, and failure-prone components. Reliability depends on validating the entire link as a whole, rather than piecing together separately procured components.
Broadcom announced in October 2025 that a test of their CPO platforms with Meta accumulated a million hours of link uptime without any “link flap”. This term describes brief connectivity losses requiring links to drop and re-establish. The results demonstrate the potential under testing conditions but do not necessarily predict behavior across all systems.
Corning and Broadcom also published a technical document on fiber design, handling, and validation practices specific to these systems. Both companies collaborate on Bailly, Broadcom’s Ethernet CPO platform with a capacity of 51.2 Tb/s. Corning supplies optical components for this system.
CPO will not replace all transceivers immediately
The advent of CPO does not mean pluggable modules will vanish soon.
Traditional optics remain suitable for many switches, servers, and connections where power, space, or electrical distance do not yet justify more complex integration.
Alternatives like linear plug-in optics (LPO) are emerging, which remove some digital processing from the module to reduce power consumption without fully moving optics into the chip package.
Each architecture balances efficiency, cost, maintenance, and interoperability differently. Pluggable modules support replacement ease and vendor competition. CPO offers greater integration and potentially higher bandwidth density but requires coordination among ASIC, optical engine, fiber, cooling, and chassis design.
Its adoption is expected first in high-capacity systems for AI networks, interconnects between accelerators, and platforms where power per bit is a critical limiting factor.
Broadcom already markets CPO platforms and has announced 51.2 Tb/s and 102.4 Tb/s Ethernet switches. NVIDIA has also revealed optical-co-packaged networking solutions for future AI infrastructure.
The real challenge is not just manufacturing fast optical engines, but producing thousands of precise, compact, and reliable links consistently and cost-effectively.
CPO shifts the fiber from the front panel into the core of the switch, reducing a key electrical obstacle for AI networks but turning the system’s interior into a complex optical environment. Scalability relies on designing and testing the entire optical route—from external connector to photonic chip—as a single, unified system.
Frequently Asked Questions
What is co-packaged optics or CPO?
It’s an architecture that places optical engines close to the network ASIC, GPU, or other processor to reduce the electrical distance signals must traverse over copper.
How does CPO differ from a pluggable transceiver?
In traditional setups, the optical module is installed on the switch front panel. With CPO, optical conversion occurs much nearer to the chip, inside the system’s main enclosure.
Why does CPO consume less energy?
Shortening the electrical route reduces signal losses and lessens the need for digital processing to compensate, saving power.
Can components of CPO be replaced individually?
Typically not, as they are not designed for field repair like pluggable transceivers. The system’s reliability and testing as a whole become more important.
via: corning
