Introduction

At the OFC 2026 technology showcase, the surge in demand for AI computing power has pushed traditional data center architectures to a critical point. Amphenol clearly pointed out in this meeting that as a single rack is transformed into a basic computing unit, interconnection technology is no longer just about data transmission, but also a comprehensive test of space utilization, power management, and thermal engineering. The core focus of this speech is on how to open up a practical and highly scalable evolution path for AI data centers between copper cable limits and CPO through innovative packaging forms such as XPO (Extra Dense Pluggable Optics) and built-in liquid cooling technology.

The physical boundaries and packaging challenges of copper cables

Despite every generation in the industry predicting the end of copper cables, Amphenol successfully pushed 224 Gbps signals to the front-end panel by directly connecting TwinX cables to the near chip interface. However, there are two major bottlenecks when moving towards the next generation. The first challenge involves breakthroughs in BGA packaging, specifically the bandwidth loss when extracting signals from ASIC packaging. The second obstacle concerns the performance of pluggable I/O, which poses a significant challenge to the electrical integrity of mechanical connectors at data rates of 224G or even higher.

In response to the uncertainty of CPO supply chain and maintainability, Amphenol has launched the XPO (Extra Dense Pluggable Optics) module, which is considered the highest density pluggable optical solution in the industry currently. The functionality of a single XPO module is equivalent to 8 OSFP modules. In a 1U space, it can accommodate 16 XPO ports, providing the same bandwidth density as CPO switches. In contrast, traditional OSFP solutions require 4U of cabinet space to achieve the same bandwidth.

(Figure 1: Comparison of XPO module density, demonstrating how a single XPO module can replace 8 OSFP modules and achieve a significant improvement in space efficiency.)

The XPO module completely abandons traditional air-cooled heat sinks. The module integrates liquid inflow and outflow pipelines internally, directly solving the thermal energy emission problem of 8 OSFP modules integrated in a very small space. This native liquid cooling integration represents a fundamental difference from traditional thermal management methods. The LPO (Linear Pluggable Optics) solution developed for XPO is expected to have a target power consumption of less than 90W, which is crucial for pursuing low latency and energy-saving AI interconnection.

(Figure 2: XPO module view, revealing the integrated liquid cooling channel and internal architecture designed for maximum thermal efficiency.)

Evolution and Expansion Challenges of Network Architecture

The transformation of data center network architecture reflects changes in the computing landscape. AI driven architecture is expanding computing networks beyond traditional boundaries. The architecture goals at the network level now have fundamental differences in three expansion dimensions: Scale Up, Scale Out, and Scale Across. Scale Up focuses on vertical in cabinet expansion, serving Fabric and storage/memory applications, with transmission distances below 10 meters, using copper cables or optical and Ethernet protocols. Scale Out processing horizontal cross cabinet connections for data center Fabric, spanning 10 to 100 meters, using optical transmission and Ethernet networks. Scale Akross enables vertical and horizontal cross cluster communication for global Fabric, covering distances of over 2 kilometers, relying entirely on optical transmission and Ethernet protocols.

(Figure 3: The network architecture displays the Scale Up, Scale Out, and Scale Across dimensions, as well as their corresponding distance requirements and transmission technologies.)

(Figure 4: Detailed table illustrating the extension architecture, applications, transmission distance, media, and technology for each extension level.)

Optimization strategy for link level and cabinet level

After surpassing the 224G data rate, there are two main implementation bottlenecks at the link level: ASIC packaged BGA breakthrough performance and pluggable I/O performance. These restrictions drive the industry towards innovative solutions that strike a balance between performance, cost, and ecosystem compatibility. At the cabinet level, CPO completely eliminates these bottlenecks by integrating optical components directly with ASICs. CPO provides superior channel reach and density compared to pluggable solutions, but introduces significant concerns about cost, supply chain complexity, and reliability.

(Figure 5: Link hierarchy architecture illustrates the main bottlenecks beyond 224G, including ASIC packaging and pluggable I/O performance limitations.)

(Figure 6: Cabinet level optimization displays the advantages of CPO (better channels, reach, density) relative to concerns (cost, supply chain, reliability).)

The industry is at a critical turning point for Co Packaged Copper (CPC), which breaks through bottlenecks by removing BGA to preserve the ecosystem while maintaining compatibility with passive copper cables or pluggable optical modules. This intermediate approach provides a migration path that strikes a balance between innovation and practical deployment considerations.

(Figure 7: CPC architecture demonstrates bottleneck removal while maintaining ecosystem compatibility with passive copper cables and pluggable optical modules.)

XPO Technical Specifications and Performance Characteristics

The XPO module introduces several groundbreaking features, positioning it as a new generation solution for AI data center requirements. The module supports 64 high-speed channels and has electrical hot swappable capability, allowing insertion and removal without shutting down the system. Operating from a single 48V power supply simplifies the power transmission infrastructure. The integrated liquid cooling system eliminates the need for traditional heat sink components, significantly reduces module height, and achieves higher port density. The XPO interface complies with the MPO-16 standard, ensuring compatibility with existing fiber infrastructure. XPO is positioned as an energy-saving alternative to DSP based solutions with a target power consumption below 90W.

(Figure 8: Density comparison shows the significant improvement in space efficiency between 16 XPO ports in a 1U space and 128 OSFP ports that require 4U.)

(Figure 9: XPO module specifications, including 64 high-speed channels, electrical hot swappable design, single 48V power supply, integrated liquid cooling, MPO-16 compatible interface, and power consumption target below 90W.)

Next generation speed targets and tuning decisions

Looking ahead to the demand for 400G per lane (448 Gbps), Amphenol pointed out that the testing validation bandwidth needs to reach 130 GHz. The industry is currently facing key decision-making points regarding the 448G modulation scheme. If PAM4 modulation is maintained, the electrical interface must maintain low insertion loss at 115-120 GHz. Adopting more complex modulation schemes such as PAM6 can reduce the requirements for physical bandwidth, but it will have a significant impact on ecosystem compatibility. This decision has a significant impact on the interconnection and testing equipment expectations of the entire supply chain.

(Figure 10: The insertion loss performance shows the progress from 10G to 224G, indicating the 200G target and future trajectory.)

(Figure 11: The insertion loss measurement updated by OFC2026 shows<2dB IL @ 53GHz, no resonance performance, and support for multiple form factors, with a bandwidth target of 448G at 130 GHz.)

The transformation plan table reveals the trade-offs involved in the next generation speed. For the "448G" target, PAM8 modulation only requires a signal rate of 150 GBaud and a Nyquist frequency of 75 GHz, while PAM6 requires 180 GBaud and 90 GHz, and PAM4 requires 112 GHz to push towards 224 GBaud. These numbers represent not only theoretical exercises, but also define the actual functional bandwidth limit at 130 GHz, and determine whether the industry can successfully expand existing infrastructure or must undergo comprehensive technological replacement.

(Figure 12: Comparison table of modulation schemes for 448G, showing PAM-N options (8, 6, 4), number of bits per symbol, signal rate, and Nyquist frequency.)

(Figure 13: Insertion loss characteristics across spectra, highlighted in the 448G region, indicating that the IL target depends on modulation, with an actual functional bandwidth of 130GHz.)