AI Photonic Interconnect Startup Lumilens Raises Over $700M in Series C

Lumilens, a photonic interconnect company connecting GPUs inside AI data centers, has emerged from stealth after raising more than $700 million in Series C funding. The financing brings its total funding to more than $900 million and values the company at $5.51 billion.

The round was co-led by Atreides Management, Bain Capital Ventures, Meritech, Seligman Ventures and Spark Capital. Addition, Alkeon, HarbourVest, J.P. Morgan Private Capital, Mayfield, Qualcomm Ventures, Peak XV, Redpoint Ventures and other investors also participated.

Lumilens said it has completed qualification of its first optical interconnect product and begun shipping into a hyperscaler’s production AI data centers. The customer was not disclosed, but the company described the agreement as being worth multiple billions of dollars.

The new capital will expand its silicon, systems and software development, process engineering and high-volume manufacturing operations.

Connectivity Becomes the New AI Bottleneck

AI data center performance is no longer determined solely by the number or speed of GPUs. Tens or hundreds of thousands of processors must exchange data quickly enough to operate as a single computing system.

AI networks can be divided into scale-up and scale-out infrastructure. Scale-up networks directly connect GPUs within a rack or tightly coupled system, while scale-out networks connect multiple racks and clusters.

Optical links are already widely used for scale-out networks, but the number of transceivers and fiber strands required increases rapidly as clusters grow. Lumilens estimates that a data center containing 400,000 GPUs could require more than 2.4 million optical transceivers and over five million fiber strands.

Scale-up networks still primarily rely on copper-based electrical connections. At higher data rates, copper has limited reach and creates additional power and thermal constraints. Connecting thousands of GPUs within a single system will increasingly require photonics, according to Lumilens.

One Platform Covering Pluggables, NPO and CPO

Lumilens targets both scale-out and scale-up networks through a common technology platform. Its scale-out portfolio includes 800G and 1.6T pluggable optical transceivers. For scale-up systems, it is developing near-package optics and co-packaged optics that move optical I/O closer to GPUs and networking chips.

The portfolio is built on LumiCore, an internally developed platform combining silicon photonics, mixed-signal integrated circuits, electrical-optical interposers and optical systems. A common platform is intended to help customers move from pluggable transceivers to NPO and CPO without replacing the underlying technology stack.

Lumilens also treats manufacturing as part of its product. It has developed proprietary process recipes, test equipment and robotics-based automation to improve production yields and accelerate high-volume manufacturing.

In May, Lumilens signed a joint development and supply agreement with POET Technologies. Lumilens placed an initial $50 million order for POET optical engines, with a roadmap extending from 800G and 1.6T transceivers to NPO and CPO. POET is a technology and manufacturing partner rather than a direct competitor.

Competition Across the Optical Interconnect Stack

Ayar Labs develops optical I/O chiplets integrated into processor and switch packages. Its TeraPHY platform moves optical connectivity closer to compute silicon to reduce electrical reach, power consumption and latency.

Lightmatter takes a broader chip-to-chip approach. Its Passage platform uses 3D-stacked silicon photonics to connect GPUs, memory and custom accelerators through a common optical layer.

Celestial AI developed its Photonic Fabric to connect processors and memory across AI systems. Following its acquisition by Marvell, that technology now strengthens Marvell’s scale-up networking and custom AI infrastructure portfolio.

AttoTude is pursuing a different solution to the same problem. Rather than replacing short-reach electrical links with photonics, it is developing a high-speed electrical interconnect designed to extend the performance and energy efficiency of copper connections.

Established optical suppliers such as Coherent and Lumentum bring large-scale manufacturing, customer relationships and broad portfolios of lasers, components and transceivers. Integrated companies including Nvidia, Broadcom and Cisco can combine networking silicon, systems and optical connectivity.

Lumilens is positioning itself between specialized photonic I/O startups and incumbent optical suppliers. Unlike companies focused on a single chip or package-level interface, it plans to cover existing scale-out networks with pluggable transceivers and future scale-up systems with NPO and CPO. Its differentiation will depend on whether its common silicon, systems and manufacturing platform can deliver multiple generations of optical products at hyperscale volumes.

Founded by a Networking Serial Entrepreneur

Lumilens was founded in 2024 by Ankur Singla, a serial networking entrepreneur. Singla previously founded software-defined networking company Contrail Systems, which Juniper Networks acquired for $176 million in 2012. He later co-founded distributed cloud platform Volterra, which F5 acquired in 2021.

The Lumilens leadership and engineering team includes veterans of Cisco, Juniper Networks, Meta, Marvell, Lumentum and Coherent.

“The constraint on AI has shifted from how many GPUs you can buy to how many you can connect,” Singla said. The company aims to supply the optical capacity required by today’s networks while enabling hyperscalers to directly connect thousands of GPUs within a single computing domain.

Lumilens has not disclosed its hyperscaler customer or the detailed delivery schedule for its multi-billion-dollar agreement. Nevertheless, qualification and production shipments within two years of founding suggest that photonic interconnects are moving from technology development into a large-scale deployment race.