DigiTimes says optical modules surge

An optical module is the plug that turns electrical signals into light so racks of graphics processors can talk to each other at speed. DigiTimes reported April 27 that those links, not just chips, are becoming the pressure point in AI data centers. (digitimes.com) (trendforce.com) The immediate shift is into faster links. DigiTimes said 1.6-terabit optical modules are nearing broad adoption in 2026, with Broadcom digital signal processors and Marvell light engines already in volume production and testing. (digitimes.com) TrendForce put a number on the ramp: AI optical transceivers are projected to grow to $26 billion in 2026 from $16.5 billion in 2025, a gain of more than 57% year over year. It said demand is rising sharply for 800G and faster modules inside AI server clusters. (trendforce.com) The constraint is not only demand. TrendForce said supplies of key laser chips remain tight, precision optical alignment still limits manufacturing scale, and power and heat are slowing deployment timelines. (trendforce.com) That is why networking vendors are trying to move optics closer to the switch chip itself. Nvidia said its Spectrum-X and Quantum-X photonics switches use co-packaged optics to deliver up to 1.6 terabits per second per port and cut power use versus traditional pluggable designs. (nvidianews.nvidia.com) (nvidia.com) Nvidia said the photonics designs can provide 3.5 times better power efficiency and 10 times higher network resiliency, while its current product pages say Ethernet photonics systems are slated for the second half of 2026. That moves the spending debate beyond graphics processors to the pipes connecting them. (nvidianews.nvidia.com) (nvidia.com) Suppliers are already expanding around that thesis. Lumentum has announced a new U.S. manufacturing facility for advanced lasers for AI data centers, and Coherent said in February it expects continued growth through fiscal 2027 on datacenter and communications demand. (lumentum.com) (coherent.com) Applied Optoelectronics has also disclosed a first volume order for 1.6T transceivers and a new volume order for 800G single-mode transceivers from a major hyperscale customer in March. Those are the parts used to move data between the thousands of processors inside large AI clusters. (investors.ao-inc.com) The result is a different map of the AI buildout. If compute remains scarce but interconnect, power and cooling are the practical bottlenecks, the next winners may be the companies selling lasers, transceivers, switch silicon and photonics packaging rather than only the processors at the center of the boom. (trendforce.com) (digitimes.com)