Lumilens Raises $900M for AI Optical Connectivity

SDxCentral reports that Lumilens' initial scale-out product completed qualification and began shipping into an unnamed hyperscaler's live AI data centers. The Silicon Boys likewise reported that the startup is supplying an optical transceiver to an unnamed hyperscale cloud provider under a multibillion-dollar commercial agreement. Lumilens has not publicly identified the customer in the retrieved coverage.

Singla told SDxCentral that the constraint on AI has shifted from GPU supply to GPU connectivity. "Whether they're training frontier models or serving them to hundreds of millions of users, hyperscalers told us they need the same two things: far more optical capacity for the networks they run today, and a path to directly connecting thousands of GPUs together into a single cluster."

The reported product areas include:

• •Pluggable optical modules for scale-out data center networks.
• •Near-package optics (NPO) for links placed near a compute package.
• •Co-packaged optics (CPO) for more tightly integrated optical and switching or compute designs.

SDxCentral reports that Lumilens designs its own silicon and characterizes LumiCore as a stack intended for high-volume manufacturing. It also reports that the company has recruited personnel from Meta, Coherent, and Cisco, including CTO Ted Schmidt, previously a senior director for silicon photonics at Lumentum.

Copper remains common for short-reach electrical links, but its power, reach, and signal-integrity constraints become more consequential as link speeds and cluster sizes increase. Optical interconnects convert electrical signals to light for transmission over fiber, enabling higher-bandwidth links over longer distances than copper connections in many data center designs.

The immediate technical distinction is important for infrastructure teams. Scale-out fabrics connect servers and racks across a cluster, where pluggable transceivers are already an established deployment model. Scale-up fabrics connect accelerators more tightly within a server or rack, where NPO and CPO seek to reduce electrical trace lengths and associated signal losses. The Silicon Boys reported that Lumilens is pursuing both layers.

Companies attempting comparable transitions from copper to optical links generally face a difficult manufacturing and systems-integration problem: component yield, thermal design, serviceability, packaging, and network operability can matter as much as raw bandwidth. The reported early hyperscaler deployment therefore provides a more concrete milestone than funding alone, although the customer, shipment volume, performance metrics, and commercial terms have not been disclosed in the retrieved reporting.

SDxCentral identifies Bain Capital Ventures, Atreides Management, and Qualcomm Ventures among Lumilens' backers. The Silicon Boys reported that the latest financing exceeded $700 million and was co-led by Atreides Management, Bain Capital Ventures, Meritech, Seligman Ventures, and Spark Capital, while total funding surpassed $900 million.

The scale of the financing places Lumilens among a growing set of companies seeking to make photonic interconnects practical for AI infrastructure. For ML platform and data center teams, the relevant question is not only whether optical links reach higher throughput, but whether they can be integrated into GPU fabrics with acceptable power, cost, reliability, and replacement procedures at fleet scale.