WeChat:
+86 13871525936
Views: 699 Author: Addams Publish Time: 2026-06-25 Origin: Site
With the commercialization of AI, data centers have become the darling of the new era. The computing power of data centers is an important bargaining chip for major manufacturers to occupy the right to speak in the AI era. At the physical layer of the data center, as the core transmission carrier of data traffic, optical modules, which were once just standard accessories for network equipment, are being pushed to the forefront of the computing power competition by AI. From the rate evolution from 400G to 1.6T, and then to the technological leap from traditional pluggable to CPO, the impact of AI on optical modules is getting wider and deeper. This article will analyze in detail the impact of the AI era on the demand for optical modules.
Before the advent of AI, 100G optical modules were the absolute mainstay of data centers, and the bandwidth of 100G was sufficient to meet business needs. However, with the advent of AI, traffic bandwidth has become the biggest obstacle limiting AI clusters. In AI high-performance data centers, the calculation volume of large models is becoming larger and larger and more complex. This is reflected in the GPU as the core of computing. In the past four years, the memory bandwidth of a single GPU has jumped from 3TB/s of H100 to 7.2TB/s of GB200, and the cluster interconnection bandwidth has skyrocketed from 400Gbit/s to 12.8Tbit/s, an increase of 32 times. This leads to the fact that if the rate of the optical module remains at 100G, it will become a de facto business bottleneck and cannot support the development of AI. In addition, AI training requires deterministic ultra-bandwidth and microsecond tail delay. These most direct needs drive the optical module rate update cycle to be compressed from the traditional 3-4 years to about 2 years. 400G has just gained a foothold, 800G will usher in large-scale sales in 2026, and 1.6T has also entered the countdown to commercial use. These real business needs are also pressuring optical module manufacturers to continuously invest costs and energy in developing optical modules with higher speeds and newer technologies. It has evolved into a new round of arms race. The winner will receive more data center orders, and those who cannot keep up will have to be slowly eliminated. This is a winner-takes-all game.
The doubling of bandwidth will inevitably lead to a further increase in power consumption, which is an inevitable price. When a 51.2T switch is fully equipped with 128 400G optical modules, the power consumption of the optical module alone may exceed 40%. In large-scale cluster deployment, the power consumption of optical modules is an important part of the data center power budget that cannot be ignored.
In high-speed optical modules, DSP is the component with the largest power consumption, but it is also indispensable. Removing it can greatly reduce the power consumption of the module. The processing of DSP has given rise to the differentiation of different technical routes.
The first is linear pluggable optical module (LPO) technology, which removes the internal DSP chip and hands over signal processing to the switch chip. It can cut the power consumption of the 400G port in half from about 10W to 5W, while reducing the cost by about 15%-20%. However, removing the DSP will lead to poor signal quality and can only be transmitted over short distances. It is the optimal solution for short-distance interconnection within 500 meters.
The second technology, co-packaged optics (CPO), is more radical. It directly packages the optical engine and the switch chip together, sacrificing pluggability and convenience, but brings the power consumption at 1.6T bandwidth to drop from 30 watts to about 9 watts, and can use water cooling to dissipate heat, which can meet the needs of different usage scenarios. At present, CPO technology is still under development, but it is not far away from commercial use.
In 2026, in the AI data center, modules will no longer be just supporting "small parts", but will determine the winner of cluster efficiency and TCO. The deployment of 400G and 800G is no longer a question of "should it", but how to do it faster, with lower power consumption and more controllable costs. Then, 1.6T will mature in 2027, and 3.2T is expected to debut in 2028. The evolution rhythm of optical interconnection has been tightly bound to the generational replacement of AI accelerators. In this computing power competition, whoever's network is ready first will realize its AI capabilities first - this will be the unchanging creed in data center construction in the next few years.
