Drive optical module demand

Moving data between AI chips is now driving a fresh buildout in optical modules, with vendors racing from 800G toward 1.6T links inside and between data centers. (lightcounting.com) (marvell.com) An optical module is the part that turns electrical signals into light so servers and switches can send data over fiber instead of copper. That matters in AI clusters because thousands of graphics processors have to exchange training data at the same time, and short, dense electrical links start to hit power and reach limits. (broadcom.com) (nvidia.com) The latest evidence is coming from suppliers, not a single headline deal. Marvell said on March 12 that its Ara 1.6T optical digital signal processor is now shipping in mass volume to global customers for 1.6T pluggable connectivity in AI data centers. (marvell.com) Credo followed on April 13 with a deal to acquire DustPhotonics, saying the purchase would bring silicon photonics technology in-house and expand its portfolio across 800G, 1.6T and 3.2T optical interconnects. (credosemi.com) Coherent used the Optical Fiber Communication Conference in March to show multiple 1.6T transceivers and early 3.2T pluggable designs. The company said the lineup covered silicon photonics, indium phosphide lasers and vertical-cavity surface-emitting lasers, or VCSELs. (coherent.com) Nvidia is also pushing the network deeper into the chip package. It said in March 2025 that its Spectrum-X and Quantum-X photonics switches would offer 1.6 terabits per second per port, with Ethernet versions due in 2026 from infrastructure vendors. (investor.nvidia.com) That shift changes where the bottleneck sits in AI systems. Nvidia says its photonics design cuts the number of lasers by fourfold and raises power efficiency by 3.5 times versus traditional approaches, while its current Spectrum-X photonics pages say the goal is to keep larger AI fabrics running longer with fewer link failures. (investor.nvidia.com) (nvidia.com) Industry forecasts show why suppliers are moving so fast. LightCounting said Ethernet optical transceiver sales doubled in 2024, rose another 70% in 2025 and could still grow 60% in 2026, though it warned shortages in accelerators and switch chips could cap deployments. (lightcounting.com) The same firm said larger AI clusters can require up to six optical transceivers per graphics processor in scale-out networks, and scale-up networks need nearly 10 times more bandwidth than scale-out. That is why optics vendors are talking about 1.6T now and demonstrating 3.2T before 800G is fully mature across the market. (lightcounting.com) (coherent.com) Telecom suppliers are seeing the same pattern outside a single building. Ciena told Fierce Network in January that it was seeing strong demand for 1.6T wavelengths, while Nokia said 400G is mainstream and 800G deployments are expected to ramp rapidly in 2026 as operators connect AI sites together. (fierce-network.com) The result is a market where the expensive part is no longer only the processor doing the math, but also the fiber links keeping those processors fed. Vendors that make transceivers, lasers, digital signal processors and photonics packaging are positioning for that spend now. (lightcounting.com) (marvell.com)