Co-Packaged Optics: Accelerating AI Data Center Networks

The rapid growth of AI workloads is straining data center networks, raising concerns about whether they can keep pace with demand. One emerging solution is co-packaged optics (CPO), which integrates optical technology directly into network switches. CPO not only enables the high speeds AI requires but also reduces power consumption—a critical consideration in modern data centers.

Figure 1. Co-Packaged Optics: Meeting the Soaring Bandwidth Needs of AI Data Centers.

Traditionally, data center switches rely on network interface cards (NICs) with transceivers that convert electronic signals to optical signals for fiber-optic transmission, using digital signal processors (DSPs). CPO streamlines this process by embedding the electronic-to-optical conversion directly into the switch’s application-specific integrated circuit (ASIC). This allows an optical cable to connect straight to the switch, delivering a continuous optical stream without intermediate conversions. Figure 1 shows Co-Packaged Optics: Meeting the Soaring Bandwidth Needs of AI Data Centers.

CPO is still in its early stages, and not all switch vendors are fully committed. However, major players are already moving forward: TSMC has developed a process for CPO-capable chips, and companies like Nvidia and Broadcom are collaborating to deploy them. Nvidia has even announced optical switches that support CPO, highlighting the technology’s real-world potential.

Why Co-Packaged Optics Is Needed

The rise of AI workloads is pushing data centers and their networks to their limits, making co-packaged optics (CPO) a timely innovation.

In the leaf-spine architecture that has dominated large data centers for over a decade, servers within a rack connect to a top-of-rack (ToR) switch via copper cables. ToR switches then connect to each other to form the data center backbone, or spine, which is typically fiber-optic. This requires each ToR switch to use an optical NIC to interface with the fiber backbone.

In AI-focused data centers, this translates to tens, hundreds, or even thousands of racks, each needing multiple NICs with optical transceivers.

Additionally, AI workloads often span multiple servers and racks, unlike traditional workloads that are more localized. While copper cables suffice for short intra-rack connections, they cannot meet the high-speed, low-latency requirements needed for longer inter-rack links. CPO addresses this challenge by integrating optical connectivity directly into switches, eliminating the need for separate NICs and transceivers.

How Co-Packaged Optics Reduces Data Center Power Consumption

AI data center networks rely on far more optical cables than traditional networks, which significantly increases power usage. Converting electronic signals to light and powering the lasers consumes energy, and every signal conversion introduces some loss, requiring extra power to maintain signal strength across multiple transitions.

“In a large AI data center, the power consumed by the optical network can be almost 10% of total compute capacity,” says Gilad Shainer, SVP of Networking at Nvidia. By removing transceivers through CPO, networking power requirements can drop by at least 3.5 times. Analyst Zeus Kerravala notes that interconnect power can be reduced by 60–70%, resulting in substantial savings per switch compared to traditional pluggable modules.

Supporting Ever-Increasing Speeds

Speed is another critical factor. As AI workloads grow, scale-up networks within racks or chassis require higher speeds, eventually reaching 400G per lane—a level where copper cables can no longer cope.

For context, an 800G module may use eight lasers at 100G each, while next-generation systems aim for 1.6 terabits per connection, with 200G per lane, eventually moving to 400G per lane. At such densities, optical connections become essential to achieve the necessary throughput.

Source: NEW ATLAS

Cite this article:

Priyadharshini S (2025), Co-Packaged Optics: Tackling the Rising Bandwidth Demands of AI Data Center Networks, AnaTechMaz, pp.180