The Future of Data Center Networking and Processing
Benoit Fleury
Published: May 8, 2025
Data center processing and networking capacities continue to be pushed to their limit with copper interconnections becoming increasingly less viable due to the rapid growth of artificial intelligence capabilities and networking models.
To meet the exponentially growing needs of data center networks while reducing power consumption created by increased processing speeds, bandwidths, and densities, optical connections must now make their way much closer to networking and processing integrated chips (ICs) to reduce the length of copper traces to these ICs. Co-Packaged Optics (CPO) achieves this by packaging the optical transceivers (often referred to as photonic chiplets) with the ICs on the same silicon substrate; this significantly reduces the length of the electrical path between optics and the electrical ICs, which in turn reduces power consumption while unlocking unprecedented levels of bandwidth concentration.
Fiber Beats Copper—and Here’s Why
The power needed to transmit electrical signals along a copper path starts to become prohibitive at data rates of 200 Gb/s, which is expected to reach the tipping point for when CPO starts to become a much better solution. The optical-to-electrical conversion that is performed by the optical transceiver is still needed in a CPO system, but it moves from a pluggable module located at the faceplate of the switching equipment to a small chip (or chiplet) that is co-packaged very closely to the target ICs inside the box. Data center chipset heavyweights Broadcom and Nvidia have both announced CPO-based data center networking products operating at 51.2 and 102.4 Tb/s.
Connecting the Dots
Connecting fibers to these transceiver chiplets, or photonic ICs (PICs), in a precisely aligned and reliable manner is a critical enabler for CPO systems. Currently, this is achieved by coupling an array of single-mode fibers to the edge or surface of the PIC. Each method has its pros and cons, and both are currently being developed by different chipset companies. The same types of fibers are used in either scenario, for which Corning is actively developing best-in-class fiber-to-chip connectivity solutions.
Early generation CPO systems, such as those announced by Broadcom and Nvidia for Ethernet switching, make use of high channel count Fiber Array Units (FAUs) that are designed to precisely align the fiber cores to their corresponding waveguides inside the PICs. These FAUs are challenging to make as they require high fiber counts, mixed single-mode (SM) and polarization maintaining (PM) fibers, integration of micro-optic components depending on the fiber-to-chip coupling mechanism, highly precise tolerance alignments, CPO-optimized fibers and multiple connector assemblies. Corning is well equipped to provide these complex inside-the-box harnesses and is ramping up operations to meet the needs of the industry as CPO technology adoption broadens.
As part of our 2025 OFC showcase, Corning launched the CPO FlexConnect™ fiber under our GlassWorks AI™ solutions —Corning’s one-stop shop for AI network needs both inside and outside the data center. The CPO FlexConnect fiber is a single-mode, bend-resilient fiber that is specifically optimized for short-length configurations, making it ideal for inside-the-box CPO fiber deployments.
The figure below illustrates an FAU assembly, with the FAU at the PIC end on the right, a SM/PM fiber harness and several optical connectors on the left that mate at the inside of the system’s faceplate.