Foundational Photonic Building Blocks Empowering Next Generation AI Infrastructure
The next frontier of AI performance demands a fundamental infrastructure shift to photonic interconnects. As agentic AI workloads create unprecedented demand for bandwidth across larger XPU clusters, the single-rack boundaries of electrical interconnects and low-bandwidth of pluggable optics have become the ultimate bottlenecks. To unlock the next-level of compute cluster performance, the industry must transition to a flatter, high-bandwidth optical fabric that blurs the lines between scale-up and scale-out networks—a shift that introduces daunting manufacturing challenges Lightmatter was founded to solve.
In this keynote, Lightmatter CEO Nick Harris will outline the photonic innovations, architectures, and roadmap required to drive 2-6 X faster AI training, prefill, and decode for the next wave of frontier models. Nick will spotlight how Lightmatter's bidirectional (BiDi) link architecture maximizes fiber bandwidth density while minimizing cabling complexity. Dr. Harris will detail Lightmatter's path to mass production, featuring COUPE Co-Packaged Optics (CPO) designs in active development with key Taiwanese partners, including GUC. Engineered for seamless integration with cutting-edge advanced packaging flows, the company’s “edgeless I/O” innovations are built on 3D photonic integration, the world's first VLSP foundry-grade lasers, and Advanced Packaging-compatible detachable fiber connectivity. This presentation delivers a definitive blueprint for how Lightmatter and Taiwan’s industry-leading ecosystem are defining and validating the future of high-yield, optical infrastructure powering the world's most advanced AI clusters.
Key Technologies Covered
- Photonic Interconnects / Optical Fabric
- Agentic AI Workloads
- AI Training and Inference Optimization
- Bidirectional (BiDi) Link Architecture
- Co-Packaged Optics
- 3D Photonic Integration
- Foundry-Grade Lasers
- Detachable Fiber Connectivity
- Edgeless I/O Architecture
- Unified Scale-Up / Scale-Out Network Architectures
