The AI era is no longer just an arena for chips, “light” is becoming the new protagonist that determines the limits of computing power.
NVIDIA (NVIDIA) recently announced on its official website that Spectrum-X, its first silicon photonic network switch using "Co-Packaged Optics" (CPO) technology, was officially unveiled and has been imported by two cloud giants, Oracle and Meta. This means that the era of AI optical communication has officially begun - the speed bottleneck of data centers is being broken by "light".
As NVIDIA enters the high-speed switching market with silicon photonics integration technology, the internal transmission structure of AI servers will be fully upgraded, driving an explosion in demand for optical modules, optical transceivers and high-speed connectors. Optical communication manufacturers in Taiwan's supply chain, such as Lianjun, Prowin, CSOT, and Shangquan, are expected to be the biggest beneficiaries of this wave.
Traditional electrical transmission is gradually unable to support the huge data throughput requirements of AI models. CPO (co-packaged optics) significantly shortens the signal transmission distance, reduces power consumption, and increases bandwidth density by directly integrating the optical module and the switching chip in the same package. Silicon photonics technology allows chips and optical components to operate on silicon substrates, forming a low-latency, high-speed, and massively scalable AI data center transmission architecture.
Is NVIDIA transforming into an "AI network architecture platform provider"?From another perspective, this means that AI hardware has entered the critical point of "converting from electricity to light." In the past ten years, the bottleneck of AI performance has been computing power, but now the bottleneck has shifted to data transmission efficiency. Whoever can take the lead in solving the problem of "internal data center latency" will have the say in the next generation of AI infrastructure. CPO is not only a technological breakthrough, but also a reconstruction of energy and architecture.
NVIDIA Spectrum-X incorporates "light" into the neural network of the AI server. It can support hundreds of terabytes of transmission per second through silicon photonic switches, and forms a "light speed interconnection" closed-loop architecture with its GPU cluster. It not only improves the efficiency of AI training, but also lays the foundation for large-scale generative AI and supercomputer computing. The industry regards this as "a new era of AI network switches", symbolizing the data center's move from "electrical interconnection" to "optical interconnection".
The author believes that NVIDIA is no longer just a chip company, but is transforming into an "AI network architecture platform provider." Spectrum-X represents NVIDIA's vertical integration of the overall data center architecture, including everything from GPUs to switches into its own ecosystem. It not only consolidates its market position, but also puts pressure on traditional Netcom giants. In the future, competition in AI infrastructure will extend from chip performance to control of the overall "network topology."
Optical communication components have become one of the core components of AI servers, and Taiwanese manufacturers have a complete optical module manufacturing chain. For example, Lianjun and Boruowei are leading the manufacturing technology of high-speed optical transceiver modules, Huaxing Optoelectronics has CPO packaging and optical engine design capabilities, and Shangquan supplies key optical connectors and high-speed transmission parts. The legal person also predicts that the demand for AI optical communications will grow at an annual rate of more than 30% in the next three years, and Taiwanese manufacturers are expected to play a pivotal role in the global supply chain.
How Taiwanese manufacturers jumped from the role of OEM to design partnerTaiwan has a solid foundation in semiconductor and optoelectronics manufacturing and has inherent advantages to take advantage of the opportunities in the CPO era. However, the challenge is that R&D and customization capabilities are still limited by the specifications dominated by major international manufacturers. If Taiwanese manufacturers can make breakthroughs in "optoelectronic integrated packaging" and "thermal management technology," they will have the opportunity to move from the role of OEM to a design partner, further enhancing their bargaining power and global visibility.
When the scale of AI training models grows with tens of billions and trillions of parameters, no matter how powerful the computing power is, it still needs "light speed" support. From GPUs to silicon photonic switches, AI infrastructure is shifting from a “computing race” to a “transmission revolution.” Whoever can control the dominance of "light" can take the lead in the AI battlefield.
This "revolution of light" is not only a technological evolution, but also an ecological shift. The next step for AI is not stronger chips, but smarter connections. From energy efficiency, heat dissipation management to module design, optical communication technology will redefine the architectural logic of AI data centers. It is foreseeable that "computing power" and "light power" will become the twin engines of AI infrastructure, driving the entire technology industry into a new round of investment and innovation cycle.
Further reading: NVIDIA leads AI optical communications into a three-year golden period! A look at the beneficiary supply chain related to silicon photonics CPO NVIDIA debuts its silicon photonic switch Spectrum-X, and the market names Taiwanese supply chain companies with eye-catching stock prices From heat to light! This new American invention "photon cooling" has greatly increased the computing efficiency of AI chips AMD silicon photonics and heterogeneous integrated R&D energy will be located in southern Taiwan Silicon photonic chips can emit light on their own, assisting the development of optical communications and CPONext：Dissatisfied with the performance in the third quarter that was not as expected, Macronix’s chairman apologized to investors and returned to the front line of production management