LightSpeed Photonics is unveiling a new near-packaged optical interconnect architecture designed to simplify the integration of optical connectivity within AI systems. The company describes the technology as a “solderable” optical interface intended to bridge the gap between pluggable optics and fully integrated co-packaged optics.
The approach places optical interconnects closer to switching and compute silicon while avoiding the complex packaging challenges associated with co-packaged optics. The architecture is designed to deliver improved power efficiency and reduced latency compared with traditional pluggable optical modules.
The company says the interconnect can be integrated directly onto system boards using standard soldering techniques, enabling higher bandwidth density while simplifying manufacturing and serviceability. The technology is being positioned for next-generation AI data center architectures where optical connectivity is moving closer to GPUs and switching silicon.
Key Points
• Near-packaged optical interconnect architecture
• Designed to reduce power and latency in AI networks
• Positioned between pluggables and co-packaged optics
• Solderable interface designed for easier manufacturing
• Targets next-generation AI cluster connectivity
“Near-packaged optics represent a practical step toward optical integration inside compute systems,” the company said. “Our approach simplifies deployment while delivering the bandwidth and efficiency required by AI workloads.”
🌐 Analysis
Near-packaged optics are increasingly seen as an intermediate step toward fully integrated co-packaged optics. While CPO promises the highest bandwidth density and lowest power consumption, it also introduces substantial manufacturing complexity.
NPO approaches attempt to place optics closer to the ASIC without fully integrating optical components into the switch package. This allows vendors to achieve improved power efficiency while maintaining the serviceability of modular optical components.
Startups exploring new optical integration approaches are emerging across the AI infrastructure ecosystem as hyperscalers search for ways to overcome the electrical interconnect limitations inside large GPU clusters.
