Optics Driving Data‑center Design
TrendForce reports NVIDIA’s architecture is accelerating adoption of co‑packaged optics in AI data centers to handle larger scale interconnects, which changes where power and latency optimizations are made (x.com). That matters for investors because it creates demand not only for GPUs but for optical components and new packaging approaches across the server supply chain (x.com).
The wires inside an artificial intelligence data center are starting to hit the same wall old highways hit at rush hour: too many lanes, too much heat, and too much wasted energy just moving traffic around. TrendForce says the next fix is not only more graphics chips from NVIDIA, but a redesign that moves light closer to the chip package itself. (trendforce.com) A modern artificial intelligence cluster is really a giant conversation between thousands of processors. Each processor trains or runs part of a model, then sends results to its neighbors, so the network becomes almost as important as the processor doing the math. (nvidia.com) Most of those links still begin as electrical signals running through copper traces and cables. Copper works well over short distances, but at higher speeds it burns more power, loses signal quality faster, and forces engineers to add extra components just to clean the signal back up. (trendforce.com) Optics solves that by turning data into light. Light can carry far more information with less loss, which is why optical links already dominate the longer connections between racks and rows inside large data centers. (broadcom.com) The catch is that traditional optical modules usually sit at the edge of a switch or server board like detachable trailers on a truck. That layout makes replacement easier, but it also means the electrical signal still has to travel across the board before it reaches the optical engine, wasting power and adding delay. (nvidia.com) Co-packaged optics changes that layout by putting the optical engine right beside the switching silicon inside the same package area. The shorter electrical path cuts signal loss, reduces the number of retimers and other helper parts, and shifts optimization work from the cable edge into the package itself. (nvidia.com) That packaging change matters more as link speeds rise from 800 gigabits to 1.6 terabits and beyond. Taiwan Semiconductor Manufacturing Company says co-packaged optics will be critical for data rates above 50 terabits per second, which is the scale large artificial intelligence fabrics are now chasing. (tsmc.com) This is the backdrop for TrendForce’s March 11, 2026 report. The firm says NVIDIA’s sixth-generation NVLink communication protocol sets 400 gigabits per second per lane as the peak transmission rate and raises the bandwidth ceiling to 3.6 terabytes per second per graphics processor, pushing data-center architects toward optical scale-up networks. (trendforce.com) TrendForce also forecasts that co-packaged optics could reach 35 percent of optical communication module share in artificial intelligence data centers by 2030. That is a big jump for a technology that spent years as a lab project and is now being pulled forward by the bandwidth demands of larger model clusters. (trendforce.com) NVIDIA has been preparing for that shift in public. In March 2025, the company said its silicon photonics switch designs could deliver 3.5 times lower power consumption, and it later said its co-packaged optics systems would offer up to 409.6 terabits per second of bandwidth with 512 ports at 800 gigabits per second, with commercial availability in 2026. (nvidia.com, nvidia.com) Broadcom is making the same bet from the switch side. In May 2025 it introduced third-generation co-packaged optics at 200 gigabits per lane, and in October 2025 it said it was shipping a 102.4-terabit co-packaged optics Ethernet switch, which suggests this is becoming an ecosystem transition rather than a one-company experiment. (broadcom.com, broadcom.com) Once optics moves into the package, the winners are no longer limited to the companies selling the main processor. Demand spreads to silicon photonics foundries, outsourced semiconductor assembly and test providers, laser suppliers, transceiver component makers, fiber-routing specialists, and thermal-management vendors that can keep hotter, denser packages stable. (trendforce.com, broadcom.com) That supply chain is already tightening upstream. TrendForce said in December 2025 that shipments of 800-gigabit-and-above optical transceivers were expected to rise from 24 million units in 2025 to nearly 63 million units in 2026, and it warned that laser light sources had become a bottleneck with lead times stretching beyond 2027. (trendforce.com) The investing angle is simple: the old artificial intelligence trade was mostly about compute chips, but the next bottleneck sits in how those chips talk to each other. If co-packaged optics moves from premium switch designs into mainstream artificial intelligence clusters, revenue should spread across optical engines, lasers, advanced packaging, and the manufacturing tools needed to assemble all of it at scale. (trendforce.com, tsmc.com, nvidia.com)
