Post in: SiliconANGLE theCUBE
The most expensive part of an AI cluster isn’t the GPUs — it’s increasingly the wire between them. In this latest interview, Ankur Singla of silicon-photonics startup Lumilens lays out why copper hits a wall past 1.6 terabits and roughly a meter, right as clusters scale toward 400,000-plus GPUs in a single data center. His answer is optics with “no technology religion” — transceivers, near-package, and co-packaged optics — manufactured at hyperscale in Singapore. For anyone tracking where the next AI margin pool forms, the interconnect is quietly becoming the story.
Watch the full interview: