NVIDIA Spectrum-X Photonics Debuts as First 200G Co-Packaged Optics for AI Factories

NVIDIA has unveiled its Spectrum-X Ethernet Photonics, marking the world’s first 200G co-packaged optics solution designed to revolutionize AI networking.

Unlike traditional optical interconnects, this new approach co-packages photonic engines directly with the switch ASIC, dramatically cutting power usage and improving performance.

The challenge NVIDIA addressed is the soaring energy and cost burden in AI factories. Compared to standard cloud data centers, AI clusters consume up to 17 times more optical power, with networking optics eating about 10% of total compute. Spectrum-X reduces that drain by minimizing the number of lasers needed and eliminating long PCB traces. For example, a 1.6 Tb/s link drops from eight lasers to just two, cutting both power and complexity while boosting reliability.

At Hot Chips 2025, NVIDIA highlighted Spectrum-X as the first implementation of 200 G/lane SerDes with silicon photonics, a standard-setting leap in electrical signaling. By deploying micro-ring modulators (MRMs) and 3D stacking between photonic and electronic layers, the solution delivers far greater bandwidth density and simpler routing. The collaboration with TSMC ensures mass production readiness for data center scale.

Efficiency gains are massive: NVIDIA claims Spectrum-X photonics brings 3.5× higher power efficiency, 10× better resiliency, and 1.3× faster deployment versus conventional optics. Its flagship Spectrum-6 102T switch integrates these technologies, replacing 64 pluggable transceivers while doubling throughput and offering 63× better signal integrity. The reduction in lasers also means longer-lasting hardware with less failure risk.

In short, NVIDIA’s silicon photonics strategy isn’t just about speed. It’s about reshaping how AI factories scale by moving energy away from wasteful optics back into the GPU clusters where real performance gains happen. With Spectrum-X, Team Green signals a future where photons, not just electrons, drive the growth of AI computing power.