Optical networking bottleneck
Analysts say AI is moving the bottleneck from raw compute to high‑speed optical interconnects, and JPMorgan has upgraded Lumentum on that thesis. As clusters scale, the cost and performance of moving data between chips and racks matters as much as the accelerators themselves. (markets.financialcontent.com)
The surprise in artificial intelligence spending is that the most valuable part of the next server may not be the chip doing the math. It may be the laser and fiber link moving data fast enough that thousands of chips do not sit idle waiting on each other. (nvidia.com) That is why JPMorgan raised its price target on Lumentum to $950 from $565 on April 9, 2026, and tied the call to demand for optical circuit switches and other data-center optics. The bet is that the traffic jam has shifted from processors to connections. (marketmoodz.com) Inside an artificial intelligence cluster, chips constantly swap pieces of a model, batches of training data, and intermediate results. If one rack can calculate faster than the next rack can receive, the expensive accelerators spend part of their time waiting. (broadcom.com) Copper cables work over short distances, but they run into power, heat, and reach limits as speeds climb and racks spread across a room. Light in fiber can carry those signals farther with less loss, which is why engineers are pushing optics deeper into the data center. (broadcom.com) The old setup used pluggable transceivers, which are small optical modules snapped into the front of a switch like removable cartridges. The new push is toward co-packaged optics, which puts the optical engine right next to the switch chip to cut power and shorten the electrical path. (nvidia.com) NVIDIA says its co-packaged optics design delivers 5 times better power efficiency and 10 times higher network resiliency than older designs built around pluggable modules. NVIDIA is building those systems for what it calls million-graphics-processing-unit artificial intelligence factories, which tells you how large the networking problem has become. (nvidia.com) Broadcom is making the same argument from the switch side. Its Tomahawk 6 chip ships with 102.4 terabits per second of switching capacity, and Broadcom says artificial intelligence clusters are scaling from tens to thousands of accelerators, turning the network into a critical bottleneck. (broadcom.com) Lumentum sits in the part of the stack that supplies the light itself. At the Optical Fiber Communication conference in March 2026, it showed high-power 1310-nanometer laser devices for co-packaged optics and 1.6-terabit transceivers for the next jump in link speeds. (investor.lumentum.com) One of Lumentum’s products is an external laser source, which is exactly what it sounds like: the light source is moved out of the hottest part of the switch and placed where it is easier to cool and replace. That matters because heat is one of the reasons optical parts become harder to scale when they are packed beside giant networking chips. (investor.lumentum.com) NVIDIA made that relationship explicit on March 2, 2026, when it announced a strategic partnership with Lumentum and said it would invest $2 billion to expand capacity and deepen research on advanced optics for data-center systems. The agreement included a multibillion-dollar purchase commitment and future capacity access rights for advanced laser components. (nvidia.com) So the JPMorgan upgrade is really a map of where artificial intelligence infrastructure spending is going next. The first wave rewarded whoever made the fastest compute chip, and the next wave is rewarding whoever can keep those chips fed with light-speed links between servers, racks, and eventually entire data centers. (marketmoodz.com)
