"Light advances, copper retreats"—a too bold move! Nvidia's new rack may feature a "combination of light and copper," with these two sectors experiencing "a tale of fire and ice."

At GTC 2026, Jensen Huang’s statement that “copper remains important, and optics will be used for different dimensional expansions; both are essential capabilities” drew market attention.

It is reported that to break through capacity limitations, NVIDIA will launch a new MGX rack—the NVIDIA Kyber. The NVIDIA Kyber is a new generation MGX NVL rack, with each rack’s NVLink domain capacity doubling, capable of housing 144 GPUs. Jensen Huang stated that this rack will utilize both CPO and copper interconnects for scale-up expansion. Additionally, the Rubin architecture is also expected to widely adopt copper interconnects.

Following this news, the concept of high-speed copper connectors became active today. As of press time, Xinya Electronics hit the daily limit, Shenyu Shares rose over 5%, while Wol Nuclear Materials, Xinke Materials, and others also gained. Meanwhile, the CPO concept sentiment was weak, with Tianfu Communication dropping over 9%, Guangku Technology and Changxin Bochuang falling more than 7%, and Robotech and Deke Li also declining.

In fact, as early as last year, Jensen Huang emphasized: “Silicon photonics still needs several years to be fully implemented. We should continue to use copper technology as much as possible. After that, if needed, we can adopt silicon photonics, but I think that will still take a few years.”

Guohai Securities pointed out that in terms of application scenarios, copper cables are more suitable for short-distance interconnects within cabinets. For example, NVIDIA’s GB200 system achieves high-speed intra-cabinet GPU interconnects via copper cables. In inference scenarios, copper cable solutions (such as AEC) are favored by cloud providers for their reliability, low power consumption, and low cost, mainly used for inter-cabinet connections in data centers, long-distance transmission, and scenarios requiring high bandwidth and low latency.

Additionally, today NVIDIA officially launched the Groq 3 LPU, which features faster memory bandwidth and can reduce latency during large model inference. Moreover, large models based on LPU not only have faster inference speeds but also offer more cost-effective pricing. Previously, Huatai Securities stated that the LPU system aims for extreme inference cost-performance, and in short-distance scale-up interconnects within cabinets, Groq emphasizes cost-effectiveness, low power consumption, and stability, thus choosing copper connections.

However, CPO high-speed optical interconnects remain the main technological evolution direction. Cinda Securities indicated that the adoption of CPO is expected to proceed in phases: in the short term, it will likely be prioritized in data center scale-out networks to alleviate the power and bandwidth pressures caused by high-speed optical interconnects; in the medium to long term, as technology matures and reliability improves, CPO is expected to achieve larger-scale applications in scale-up high-speed interconnects.

During the GTC 2026 keynote speech, Jensen Huang showcased the Spectrum-6 Ethernet switch, an important component of the AI supercomputing platform Vera Rubin, which uses CPO technology to deliver five times higher optical power efficiency and ten times higher network reliability.

(Source: Cailian Press)