In April 2026, when Nvidia's GPU cluster was training trillion-parameter large models day and night, an invisible bottleneck was breaking out: the chips were too slow to "speak".

Just like Beijing's Fifth Ring Road in the morning rush hour - no matter how many cars there are, the road is not wide enough, and they all have to be blocked.

Optical communication is an ultra-wide, low-latency, and zero-congestion data highway tailored for AI data centers.

From "as long as it can be used" to "it must be"

At the Global Optical Communication Conference (OFC) in March 2026, 1.6T optical modules became the focus of the audience. It has been found that the industry has accelerated from 800G to 1.6T - Lumentum even announced that it will start mass production of 1.6T optical modules this summer. Although AI data centers will still mainly use 800G in 2025, 2026 will already be the "first year of pre-mass production" of 1.6T.

Why is the upgrade so urgent?

Because today's AI training often uses thousands or even tens of thousands of chips to work together. In the past, the method of connecting with copper wires has hit the physical ceiling in terms of power consumption, heat generation and transmission distance, and it cannot be carried at all.

As a result, optical interconnection technology—that is, using optical signals instead of electrical signals to transmit data—is no longer a "dispensable" option, but a must-use solution.

An industry analyst gave an analogy: the previous data center was like a country road, and copper cables could still cope; Today's AI data centers are like city centers in the morning and evening rush hours, and they are completely blocked without a "highway".

Optical communication is this information highway.

Lumentum's layout reflects this transformation. It not only makes ordinary optical modules, but also focuses on the core optical devices required by AI data centers: high-power lasers, EML (electrical absorption modulation lasers), optical path switching equipment (OCS), and providing external laser light sources for "co-packaged optics" (CPO).

These technologies are not prepared for ordinary data centers, but are specially tailored for the CPO architecture of NVIDIA GPU clusters - simply put, to "build highways" for the strongest AI computing power.

Why is the demand for optical communication suddenly so strong?

Lumentum's orders are scheduled until 2028 for three key reasons:

The AI "factory" is getting bigger and bigger, and communication must be faster.

Today's AI training no longer relies on a single chip, but uses thousands of chips to form superclusters. But the more chips, the greater the pressure to transmit data to each other - just like in a large factory, if workers cannot communicate quickly with each other, even if everyone is capable, the overall efficiency will be stuck

Optical communication is the "internal high-speed network" that allows these "chip workers" to have real-time and efficient conversations. Without it, no matter how strong the computing power is, it cannot be exerted.

Giants no longer "buy ready-made", but pull suppliers to build together.

In the past, companies like Nvidia, Google, and Microsoft would directly purchase standard optical modules.

But now, they regard optical communication as the core part of the AI system and are deeply bound to suppliers: joint research and development, joint design, and customized integration

For example, Nvidia and Lumentum have signed a multi-year strategic cooperation agreement, not only to buy products, but also to engage in technology development together. Lumentum is no longer just a "parts supplier", but has become a "technical partner" of AI giants. This relationship is closer, and orders are more stable and larger

Technology is rapidly upgrading: from 800G to 1.6T.

2025: The mainstream is still 800G optical modules 2026: 1.6T will start mass production and enter the "climbing period" But 1.6T is not a simple speed-up, it requires new materials, new processes, and new testing methods, and the technical threshold is very high. Whoever can break through first will be able to eat the biggest piece of cake in the next few years

Chinese companies are not idle either

When Lumentum's orders have been scheduled for 2028, Chinese companies are not idle and are quickly seizing the exploding market. Companies such as Zhongji Innolight, Xinyisheng, and Huagong Technology have chosen different technical routes to accelerate their layout.

Zhongji Innolight: The "leader" of global optical modules

It is currently one of the most important optical module suppliers in the world. It is expected to account for 35% of the global 800G optical module market by 2026, and its main customers include Nvidia, Google and other North American manufacturers

It is also at the forefront of 1.6T optical modules, and large-scale mass production has been achieved in the second quarter of 2026. A recent specific example: on April 7, 2026, Google placed a large order for 12 million NPO (near-packaged optics) optical modules. Zhongji Innolight won 60% (7.2 million) and Xinyisheng got 40% (4.8 million), with a total order value of about 12 billion to 15 billion yuan

Xinyisheng: Take the differentiated route of "saving electricity and money"

Instead of completely following the mainstream, Xinyisheng bets on LPO technology (linear drive pluggable optics).

Compared with traditional solutions, LPO has lower power consumption and lower cost, making it especially suitable for high-speed connections over short and medium distances in AI data centers. It is expected that in 2026, its 800G LPO module shipments will reach 3 million

It's not who replaces whom, but each has its own division of labor

Lumentum stands at the top of the industrial chain, mastering high-barrier technologies such as core lasers; Chinese companies have advantages in large-scale manufacturing, rapid delivery and application innovation

This is not a "you die and I live" replacement war, but a new pattern that is both competitive and complementary - the explosion of global AI computing power is large enough for different characters to find their place

Jingtai View|Three main lines, grasping the "AI infrastructure water seller"

Main line 1: Global leader + technology card holder

Lumentum (U.S. stock): upstream device monopoly, CPO/OCS two-wheel drive; Coherent (U.S. stock): silicon light + high-power laser, NVIDIA's strategic partner

Main line 2: Made in China + large-scale going overseas

Zhongji Innolight: The world's main supplier of 800G/1.6T, deeply bound to the North American cloud factory; Xinyisheng: Differentiated LPO route, significant cost advantage, and large growth elasticity

Main line 3: Domestic alternative potential stocks (observation period)

Huagong Technology, Guangxun Technology, Yuanjie Technology, etc. are deployed in the field of EML chips and silicon optics; it is difficult to shake overseas leaders in the short term, but in the long run, it is independent and controllable

Risk warning: rapid technological iteration, geopolitical disturbances, and overcapacity concerns.