Share with your CTO
Lumilens is betting that the binding constraint on AI at scale has already shifted from GPU count to GPU connectivity, and $700 million from a broad syndicate including Bain Capital Ventures, Atreides Management, and Spark Capital suggests serious investors agree. The US-based optical interconnect startup emerged from stealth at a $5.51 billion valuation, two years after founding. CEO Ankur Singla previously built and sold two infrastructure companies to Juniper Networks and F5. The capital goes toward silicon photonics manufacturing and high-volume production of the optics needed to link thousands of GPUs into a single coherent cluster.
What this means for your business
Whether Lumilens is on your radar depends less on your vendor shortlist today and more on how your AI infrastructure roadmap handles scale. If your organization is running GPU clusters in the dozens or low hundreds, electrical interconnects still work fine. Once you’re planning for thousands of GPUs operating as a unified system, the physics of copper wiring become the ceiling, and optical interconnects, which use light instead of electrical signals to move data between chips, become the floor your architecture rests on.
Singla’s framing, that capacity is abundant and connectivity is scarce, is worth stress-testing against your own hyperscaler relationships. The major cloud providers are already building their own optical networking capabilities, and companies like Arista, Cisco, and several well-funded startups are competing for the same layer of the stack. Lumilens is specifically targeting co-packaged optics, where the optical components sit directly alongside the chip rather than at the edge of the rack, which reduces latency and power draw but demands tighter integration with the chip itself. That’s a manufacturing and partnership problem as much as an engineering one, and $700 million buys runway, not certainty.
The genuine signal here isn’t the valuation, it’s the founding team’s track record. Singla has twice built infrastructure companies that large networking incumbents found worth acquiring. That pattern suggests he’s reading where the chokepoints form before the incumbents do. If you’re in the middle of a multi-year GPU cluster procurement or an AI data center buildout, the interconnect layer deserves explicit attention in your architecture reviews now, not when the copper bottleneck shows up in your performance telemetry. I’d revise this view if Lumilens’ manufacturing ramp stalls and hyperscalers accelerate in-house optical development instead of sourcing externally.
Concept deep-dive: Co-packaged optics
Traditional optical transceivers plug into the edge of a network switch, the way a USB drive plugs into a laptop port. Co-packaged optics move the optical components directly onto the same package as the switch chip, the way RAM is soldered onto a motherboard rather than inserted into a slot. The result is lower power consumption and higher data throughput at short distances. For AI clusters connecting thousands of GPUs, those gains compound into the difference between a system that scales and one that saturates.
Based on reporting from Peak XV, EDBI join $700m funding round of AI startup Lumilens, originally published 2026-08-06 23:29:00.
