Insider Brief
- Lumilens emerged from stealth with more than $900 million raised, including a fresh $700 million-plus Series C that values the AI data-center networking startup at $5.51 billion.
- Atreides Management, Bain Capital Ventures, Meritech, Seligman Ventures and Spark Capital co-led the round, with participation from Addition, Alkeon, HarbourVest, J.P. Morgan Private Capital, Mayfield, MVP Ventures, Qualcomm Ventures, Peak XV, Redpoint Ventures and Seifdune.
- The company is already shipping its first product under a multibillion-dollar agreement with an unnamed hyperscaler and is developing 800G and 1.6T pluggable transceivers for scale-out networks along with NPO and CPO systems designed to extend scale-up connections beyond copper limits.
Lumilens has announced emerging from stealth with $700 million in Series C funding, bringing total funding to $900 million, as the two-year-old startup expands optical networking technology designed to connect the growing numbers of processors used in AI data centers.
Atreides Management, Bain Capital Ventures, Meritech, Seligman Ventures and Spark Capital co-led the Series C round, according to the San Jose-based company. Addition, Alkeon, HarbourVest, J.P. Morgan Private Capital, Mayfield, MVP Ventures, Qualcomm Ventures, Peak XV, Redpoint Ventures and Seifdune were among the other investors. The company said the new funding brings the company’s valuation to $5.51 billion.
Lumilens indicated its first product is already shipping into an unnamed “hyperscaler’s production AI data centers” under what it described as a multibillion-dollar agreement.
Lumilens is targeting networking constraints that are becoming more important as AI data centers grow to hundreds of thousands of graphics processors. Those processors must exchange large amounts of data quickly enough to operate as a coordinated computing system, making the connections between chips an increasingly important part of overall performance.
“The constraint on AI has shifted from how many GPUs you can buy to how many you can connect,” founder and CEO Ankur Singla, said in the announcement. “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. We built Lumilens to deliver both — and to manufacture at the volume, quality, and pace this buildout demands.”
LumiCore Platform
Lumilens is developing products for both major networking layers inside an AI data center, covering what the company describes as the full connectivity stack.
Both product lines are built on Lumilens’ LumiCore platform, which combines silicon photonics, mixed-signal integrated circuits, electrical-optical interposers and optical systems developed by the company. Lumilens said using a common technology platform allows it to develop products for both scale-up and scale-out networks while designing them for high-volume manufacturing.
Scale-out
The first problem comes from the networks connecting racks and rows of GPUs. As AI clusters grow, each additional processor increases the number of optical transceivers and fiber connections required. Lumilens said a 400,000-GPU data center could need more than 2.4 million transceivers and five million fiber strands. It pointed out that McKinsey projects that production of 800-gigabit transceivers could fall 40% to 60% short of demand through 2027, with 1.6-terabit shortages of 30% to 40% continuing through 2029.
Lumilens indicated it is targeting that problem with pluggable optical transceivers supporting 800G, 1.6T and higher speeds, designed to move data between racks and rows over fiber.
Scale-up
A different constraint exists inside the rack, where GPUs are connected directly so they can operate as one computing system, the company noted. At the data rates required by large AI workloads, Lumilens pointed out that electrical signals can travel only about 1.5 meters over copper, effectively limiting tightly connected groups to a single rack and hundreds of processors.
Lumilens said it is addressing that limit with near-package optics and co-packaged optics, known as NPO and CPO, which bring optical connections directly to or near the GPU. The technology is designed to connect thousands of processors as a single tightly coupled system, with tens of thousands on its longer-term vision.
