NVIDIA's CPO switches have arrived, marking a significant upgrade in networking infrastructure.

21st Century Business reporter Niyi Qian reporting from San Jose

The rapid development of AI computing has put higher requirements on the networking infrastructure of data centers. On March 18, NVIDIA launched its new NVIDIA Photonics silicon photonics technology at the GTC2025 conference.

According to the introduction, this technology replaces traditional pluggable optical transceivers with co-packaged optics (CPO), allowing fibers to be directly connected to switches, significantly reducing the power consumption of data centers. NVIDIA estimates that this technology can reduce power consumption by 40MW and improve the network transmission efficiency of AI computing clusters, laying the foundation for future large-scale AI data centers.

Based on this, NVIDIA launched the Spectrum-X and Quantum-X silicon photonics network switches, integrating electronic circuits and optical communication technologies deeply, helping AI factories connect millions of GPU clusters across regions to reduce energy consumption and operating costs.

The Spectrum-X Ethernet platform is designed for multi-tenant large-scale AI factories, with a bandwidth density 1.6 times that of traditional Ethernet, supporting the world's largest supercomputer. The Quantum-X photonic InfiniBand platform provides 144 800Gbps ports based on 200Gbps SerDes technology, adopts liquid cooling design to efficiently cool silicon photonics modules, and speeds up AI computing architecture by 2 times and expands by 5 times compared with the previous generation.

As AI factories expand on an unprecedented scale, networking technologies urgently need to be innovated. NVIDIA said that its photonic switches integrate innovative optical communication technologies, reducing the use of lasers by 75% compared with traditional methods, increasing energy efficiency by 3.5 times, enhancing signal integrity by 63 times, and improving network resilience by 10 times, and deploying speeds are 1.3 times faster.

NVIDIA founder and CEO Jensen Huang said: "AI factories are a new paradigm for large-scale data centers, and the networking infrastructure needs to be completely restructured. By directly integrating silicon photonics technology into switches, NVIDIA has broken the limitations of traditional large-scale and enterprise networks and paved the way for AI factories with millions of GPUs."

NVIDIA's silicon photonics technology ecosystem covers TSMC, Browave, Coherent, Corning, Fabrinet, Foxconn, Lumentum, SENKO, SPIL, Sumitomo Electric and TFC Communication.

Guoyuan Securities' report previously predicted that NVIDIA would launch a new CPO (co-packaged optics) switch at the March GTC conference. This CPO switch is in the trial production stage, and if everything goes well, it can be mass-produced in August this year.