Lumilens came out of stealth on Thursday with a big number and a bigger claim. It announced more than $700m in fresh funding, taking its total to over $900m and valuing it at $5.51bn. It also says it is already shipping its first product into a hyperscaler’s live data centres, under a deal it says is worth billions.
The pitch rests on a shift the whole industry is feeling. “The constraint on AI has shifted from how many GPUs you can buy to how many you can connect,” said founder and chief executive Ankur Singla. Modern models run across hundreds of thousands of chips that have to behave as one computer. The wiring between them now matters as much as the chips.
Copper hit a wall
Two problems bite. Across racks, every GPU you add multiplies the optical transceivers needed. A single 400,000-GPU site needs more than 2.4 million of them, and the market cannot keep up. Inside the rack, GPUs are wired to each other over copper, and copper is out of road.
At AI speeds an electrical signal survives about a metre and a half over copper. That caps a tightly linked cluster at a few hundred chips. To wire thousands together, you have to swap copper for light. That is Lumilens’s business: optical interconnects for both jobs, the short hops inside a rack and the long ones between them. It is the same copper bottleneck that pushed Nvidia into photonics.
