$NVDA $AVGO $TSM $LITE $COHR Schulz Class #11#: co-packaged optics, or moving the light right up to the switch chip Class #10# ended by taking the silicon photonics chip out of the plug-in module. Here's why that's worth doing. A 1.6T plug-in optical module burns about 30 watts, and NVIDIA says more than half of that goes to one chip whose job is to clean up the signal after a 14-to-16-inch trip across the circuit board. Co-packaged optics cuts that trip to under half an inch. What it is Think of a restaurant with the kitchen in the middle of the building and the takeout windows along the front wall. Every order gets carried down a long corridor, and each window keeps its own staff to fix whatever got mixed up on the way. Co-packaged optics moves the windows right next to the stove. In an AI network switch, the kitchen is the switch chip and the windows are the plug-in optical modules from Class #7#. The electrical signal crosses the board to reach them, and each module carries a DSP (digital signal processor) to repair it. Co-packaged optics, or CPO, puts the silicon photonics chips on the same package as the switch chip, so the signal turns into light almost right away and the DSP isn't needed. The lasers stay outside. NVIDIA puts them in replaceable modules on the front panel, because they're the parts most likely to fail. Why it matters NVIDIA's Spectrum-X Ethernet Photonics switch, built on CPO, has been in full production since early June, and NVIDIA says it's 5 times as power-efficient as a network built on traditional plug-in modules. Broadcom's third-generation CPO switch, Tomahawk 6 Davisson, moves 102.4 terabits per second and uses 70% less power for its optical links than plug-in modules do. In a data center that's short on power, every watt the network saves can go to GPUs instead. TrendForce said in July that CPO had officially entered mass production, with optical engine yield and advanced packaging capacity as the bottlenecks. What it means for stocks NVIDIA and Broadcom sell the switches. TSMC makes the optical engines for both, on its COUPE platform. NVIDIA names Lumentum, Coherent and Sumitomo as suppliers of its external laser modules, so each switch still runs on the indium phosphide lasers from Class #9#, with each laser shared across several lanes. That wraps up the optics half of the course. Every one of these engines has to be known good before it's sealed next to a switch chip, so the next four classes are about testing. A few weeks ago we wrote about how much engine capacity NVIDIA has lined up for these switches. What would you like the next class to explain?