The Future of AI Data Center Networking: CPO, XPO and Optical Innovation

Riding the wave from the 2026 Optical Fiber Communication Conference (OFC) in March, there has been growing discussion around co-packaged optics (CPO) and a broader ecosystem of technologies shaping the future of AI data center networking. To understand why this shift is accelerating, it helps to look at how data center connectivity has evolved and what AI is changing now.

The Evolution of Data Center Connectivity
AOI was involved in data center optics early on when hyperscale operators first began adopting optical connectivity inside their facilities. Before that shift, most connections inside a data center relied on copper. That worked when data centers were physically smaller, and the speeds required between servers and switches were lower.

As data centers grew and networking speeds continued to climb, copper stopped being an effective solution over the distances involved. That’s when optics started moving inside the data center. AOI was an early innovator in this transition, developing optical transceivers specifically for this environment.

Why AI Workloads are Driving Higher Bandwidth
Today the industry is going through another major transition, largely driven by AI. Training large AI models requires large compute clusters—often thousands of GPUs or accelerators working together. Those clusters depend heavily on high-speed connectivity between nodes, which in turn requires a rapid increase in bandwidth inside the data center.

This can be seen in the transition of optical speeds. 400G optics are now widely deployed. While 800G optics have become the standard for modern AI clusters, the industry is now aggressively deploying 1.6T connectivity to keep pace with the massive input and output requirements of next-generation GPU architectures.

To manage the thermal and power challenges of these speeds, the industry is gravitating toward Linear-Drive Pluggable Optics (LPO) and Linear-Receive Optics (LRO). By removing or minimizing the DSP within the module, these technologies offer a path to significantly lower latency and power consumption while preserving the operational flexibility of the pluggable form factor.

What is Co-packaged Optics (CPO)?
As those speeds increase, system designers start running into new challenges. One of the biggest challenges is the electrical path between switching silicon and the optical modules themselves. As data rates get higher, maintaining signal integrity across those electrical interfaces becomes increasingly difficult. That’s where CPO comes in.

The basic idea behind CPO is to co-locate the optical engines and the switching ASIC on a single substrate. This transition is supported by innovations like External Laser Sources (ELS), which move the sensitive laser components to the front panel for better thermal management and reliability, while keeping the high-speed modulation inches away from the silicon.
However, CPO is not a standalone solution. It is part of a broader ecosystem of complementary technologies that define the future of AI networking.

The AI Optical Ecosystem
Rather than a single transition, the industry is evolving toward a layered architecture, where different standards address distinct parts of the AI networking stack. These include:

400G Optical: The Scaling Foundation - This establishes the baseline for AI backend networks by doubling interface speeds to 400G per lane. It leverages IEEE methodologies and enhanced FEC to support the density requirements of gigawatt-scale clusters.
eXtra-dense Pluggable Optics (XPO): Pushing Front-Panel Density - XPO extends the pluggable model to 12.8T per module. To manage this density, it introduces "belly-to-belly" PCB designs, high-leverage mechanical ejectors, and integrated liquid cooling to support 400W+ power loads.
Open CPX: Standardizing Co-Packaged Optical Engines - Open CPX creates a "Lego-like" ecosystem for co-packaged optics by standardizing sockets and connectors. This modular approach enables a multi-vendor market for interoperable optical engines, targeting annual volumes of 100M+ ports.
Optical Compute Interconnect (OCI): Ultra-Low Power Scale-Up Fabrics – Targeting chip-to-chip communication where copper fails, OCI achieves extreme efficiency. It uses 4-wavelength DWDM and bidirectional fiber signaling to enable high-speed, low-latency fabrics for GPU clusters.
6.4T On-Board Optics featuring ELSFP Lasers: At OFC 2026 in Los Angeles, AOI showcased its 6.4T OBO (On-Board Optics) technology powered by AOI's high-power ELSFP. The OBO architecture enables higher aggregate bandwidth and superior power efficiency, providing an immediate, high-density alternative to NPO and CPO architectures specifically optimized for the rigorous signal integrity requirements of hyperscale AI infrastructure.

Supply Chain Considerations for Optical Components
This shift redefines how optical systems are designed, manufactured, and integrated. Many of the optical architectures being discussed today rely on larger devices that require more specialized semiconductor manufacturing processes. As a result, they consume more wafer real estate and require more manufacturing capacity.

That’s one reason the industry is paying closer attention to optical component supply chains. In particular, indium phosphide fabrication (InP) capacity, which is used to manufacture many of these lasers, is becoming a critical part of the equation. AOI’s history is unique in that respect. The company was originally founded around laser technology and has been manufacturing optical devices internally for many years. Having that capability in-house allows AOI to expand capacity as demand grows and gives customers confidence that the key components needed for these systems will be available.

Scaling Optical Manufacturing for AI Infrastructure
Manufacturing at scale is the other important part of this conversation. The volume of optical connectivity required in AI infrastructure is growing very quickly. That means the industry needs to be able to ramp production of optical modules in a predictable way. Over the past several years AOI has invested in automated manufacturing for optical modules, and more recently expanded our domestic manufacturing footprint. Automation allows the replication of production lines more easily and the scaling of capacity as demand increases. It also provides flexibility in where that production takes place, which is becoming more important as hyperscale operators pay closer attention to supply chain security.

The industry isn’t moving from one architecture to another overnight. A set of technologies—pluggable optics, XPO, CPO, and emerging chip-level interconnects—are evolving together to support the rapid growth of AI infrastructure. As speeds continue to increase and compute clusters continue to scale, layered approaches are likely to become an important part of the industry’s longer-term roadmap.