Balerion Principal Aidan Daoussis sits down with Vivek Raghunathan, Co-Founder, CEO, and CTO of Xscape Photonics, to discuss optical interconnects for AI. Xscape is developing multicolor laser technology to increase communication bandwidth between GPUs, addressing data movement bottlenecks in memory-intensive AI workloads. The conversation covers silicon photonics, data center power constraints, and the manufacturing economics required for hyperscale deployment.
00:00 – Introduction to Xscape Photonics and AI hardware efficiency
04:36 – GPU memory access, escape bandwidth, and communication bottlenecks
07:16 – Using multicolor lasers to increase bandwidth between GPUs
08:07 – Raghunathan’s background: MIT, Intel, and commercializing silicon photonics
15:47 – Developing co-packaged optics at Broadcom
18:26 – Xscape’s origins and wavelength division multiplexing
20:34 – Laser modules, strategic investors, and wafer-scale manufacturing
21:40 – Data movement power consumption and GPU utilization
25:48 – Data center growth, grid constraints, and inference efficiency
30:41 – AI agents, productivity, and the purpose of expanding compute infrastructure
33:58 – How AI could change data center architecture38:20 – Moving from laboratory demonstrations to commercial deployment
42:18 – Why AI demand changes the opportunity for photonics
47:33 – Investor diligence, technology differentiation, and competitive risks
50:31 – Closing perspective on optics and future AI infrastructure
