Lumilens

Lumilens develops photonic interconnect products aimed at the networking bottlenecks that arise as AI clusters grow. The company frames the problem as expanding model sizes and inference volumes driving demand for larger GPU clusters, which pushes existing scale-out architectures against physical, electrical and thermal limits and requires new approaches to scale-up (GPU-to-GPU) architectures. Its stated positioning is to lead in both scale-out and scale-up optical connectivity.

The product line is built on the LumiCore platform, described as a fully integrated silicon platform combining silicon photonics, mixed-signal ICs, electrical-optical interposers and optical systems. A single platform and software architecture supports pluggable optics, near-package optics (NPO) and co-packaged optics (CPO), allowing hyperscalers to begin scale-up deployments with NPO and transition to full CPO on the same silicon platform. Stated performance claims include scale-up connectivity extending beyond 100x current GPU direct-connect limits with minimal latency, and a 1000x bit error rate improvement for scale-out that reduces retransmissions.

Alongside device and IC design, Lumilens develops its own manufacturing capability, including process recipes, test equipment design, automation and MES, with end-to-end robotic automation and AI-optimized processes intended to raise yields and allow production in multiple locations. The company says initial products were designed, qualified and shipped in under two years, and that multiple patents are pending.

Founding story

Lumilens was founded to address the gap between the pace of AI infrastructure growth and what the photonics industry could deliver, starting from what its founders describe as a clean sheet of paper to handle growing AI cluster sizes. Founders include CEO Ankur Singla, VP Products Samuel Liu, VP/GM India Ritesh Kapahi, VP Operations Dave Friedman and CTO Ted Schmidt.

Business model

Lumilens designs and manufactures optical interconnect hardware — pluggable optics, near-package optics and co-packaged optics built on its LumiCore silicon platform — and sells it to hyperscale AI data center operators, including under long-term customer agreements. It retains in-house design of silicon photonics and mixed-signal ICs as well as ownership of manufacturing processes, test equipment and automation, and offers customization of scale-up architectures to individual hyperscaler roadmaps.

Traction

Lumilens states it is shipping optical interconnect products into production hyperscaler AI data centers under a multi-billion-dollar customer agreement, that its initial products were designed, qualified and shipped in under two years, and that it has more than 100 staff.

Latest developments

The company emerged publicly announcing more than $900 million in funding and reporting production shipments into hyperscaler AI data centers under a multi-billion-dollar customer agreement. Press coverage reported a $700 million raise at a $5.5 billion valuation ($5.51 billion), and executives have appeared in public discussions on scaling AI factories and large data centers.