Marvell plans $115 billion acquisition of Lumentum, China's optical communications faces new test

On November 2, Marvell announced a proposed all-stock acquisition of optical communications vendor Lumentum, a transaction corresponding to an enterprise value of approximately $115 billion. That figure corresponds to the entire company, and of course cannot simply be understood as the cost of "speeding up" the network between chips. But why a chip company would be willing to strengthen its optical communications capabilities through such a large-scale transaction is still worth asking.

When we talked about AI infrastructure in the past, we were used to looking at what GPUs were used and how many were installed. This transaction pushes another question to the fore: if you buy a lot of chips, can they actually get more work done? Following this question downward, the opportunities and changes facing Chinese optical communications companies will also become somewhat clearer.

1. Buying the chips back does not mean computing power is fully utilized

Imagine a factory where the boss buys a batch of extremely fast machines, but the conveyor belt cannot keep up. Some machines have finished, while others have not yet received materials; the equipment is all advanced, yet the entire production line may not be efficient. AI clusters encounter similar problems: strong performance in a single chip does not mean that, put together, they can stack up to equally strong capability according to their number.

Training large models usually requires a large number of chips to divide the work and cooperate. Each completes part of the computation, and they also need to exchange data, aggregate results, and then continue to the next step. If the connections cannot keep up, the chips may be waiting—not because they cannot compute, but because the data needed for the next step of computation has not yet arrived. Waiting cannot necessarily be eliminated entirely, but if it takes up too much time, expensive equipment is not being fully utilized.

Therefore, to look at an AI cluster, one cannot merely count how many GPUs are installed; one must also see how well they work together. Whether the network transmits fast enough, whether there is congestion, and whether data transmission can proceed simultaneously with computation all affect how quickly the entire system completes tasks.

How many chips are bought is input; how much work these chips ultimately complete is output.

Of course, not all AI systems are bottlenecked on the network. Different tasks have different constraints, and memory, power, cooling, and software can all affect efficiency. Connectivity receives more attention because chips are becoming more powerful and clusters are becoming larger, and the network responsible for making them cooperate must also keep up. Otherwise, improvements in single-chip performance may not fully translate into efficiency improvements for the entire system.

2. Fiber has long been in use, so where is the new change?

When people see "optical communications," some think of fiber broadband, and others think of "optical computing." These two things must be distinguished: the optical interconnect discussed in this article mainly converts data into optical signals for transmission, then converts them back at the other end. It solves the problem of data transmission, not letting light replace GPUs to complete computation.

Data centers have long been using optical fiber, and copper connections still have applicable scenarios. The change worth watching in this round is that optical connections are moving closer to chips. In common pluggable optical module solutions, the electrical signal emitted by the switch chip must first pass through a stretch of circuit board, reach the optical module, and then be converted into an optical signal. As transmission rates increase, how to control the signal loss and power consumption of this electrical connection becomes more challenging.

One approach that has received attention is called CPO, which in Chinese is "co-packaged optics." It can be understood as moving the optical engine responsible for optoelectronic conversion next to devices such as switch chips and placing it in the same package. The path that high-speed electrical signals need to travel becomes shorter, creating an opportunity to reduce the loss and power consumption pressure of this connection.

But being closer also means greater mutual influence. How heat is dissipated, how optical fibers are connected, how to test after manufacturing is complete, and how to maintain in the event of failure all need to be considered together during design. What customers ultimately need to compare is the performance, reliability, and total cost of the entire solution over its service life, not just how much power a single component saves.

This change is not only an upgrade in transmission speed; it may also change what kind of products customers buy. Components that were relatively independent in the past will be more tightly combined in some new solutions. This pushes the question from "who can provide a good part" to "how these parts can work well together from the very beginning."

3. What Marvell wants to buy is not just more parts

This is also one angle for understanding the combination of Marvell and Lumentum. Marvell has technologies such as interconnect, network switching, and custom chips, while Lumentum has capabilities such as lasers and optical transceivers. In the announcement, the two sides proposed that they hope to strengthen the integration and co-design of the electrical and optical signal chain.

Parts being excellent individually does not mean they will achieve the best effect when assembled together. How a chip sends signals affects the requirements of subsequent components; where the optical engine is placed affects packaging and heat dissipation. The tighter the product integration, the harder it is to wait for each party to finish its part and then rely on the final assembly step to solve all problems.

English text is being prepared.

However, combining the companies does not mean the products will naturally work well together. Whether R&D teams can collaborate, whether solutions can be launched on time, and whether customers are willing to purchase all have to be proven through actual delivery. The transaction is expected to be completed in the first half of 2028 and still requires approval from both sides' shareholders and regulators.

The transaction amount shows the weight of the bet, but it cannot prove in advance that the bet is correct. What ultimately determines value is still whether the integrated products can solve customers' problems and whether customers are willing to continue buying.

4. The opportunity for Chinese companies is not just selling more products

There may be more places where optics are used, but what customers want to buy may have changed. This is also a new test that Chinese optical communications companies need to face.

Chinese companies are already involved in high-speed optical connectivity and are also preparing next-generation products. Zhongji Innolight's products include 800G and 1.6T high-speed optical modules; TFC Optical Communication disclosed volume production of 1.6T optical engines and R&D progress on CPO-matched optical components. They operate at different stages of the chain, but both are related to upgrades in connectivity technology.

This does not mean CPO will quickly replace all optical modules. Under different distances, different equipment and different cost requirements, multiple connectivity approaches will still coexist. But where new integration solutions are adopted, some functions will move closer to the chip, and product boundaries may also adjust accordingly: things originally sold as standalone modules may enter systems in different forms in the future.

The functions are still needed, but customers' purchasing and validation methods may change. Companies must both seize demand for existing products and judge what they need to provide for next-generation solutions.

Existing capabilities will not lose their value overnight because of this. Light sources need to be supplied, optical fibers need precision connections, and complex devices need reliable packaging and testing. Being able to manufacture these things stably and control cost, quality and delivery well is itself competitiveness. The key is whether, when customers change to a different design, these capabilities can follow into new products.

Therefore, large-scale M&A by overseas giants can serve as a clue for observing industrial changes, but it cannot be directly treated as evidence that Chinese manufacturers will gain more orders. For a specific company, R&D completion, sample delivery, passing customer validation, mass production and generating revenue still need to be viewed separately. Between making a new product and delivering it stably and in large volumes to customers, there are often still many engineering problems to solve.

Conclusion

For Chinese optical communications companies, there are actually two things to do right now: on one hand, do existing products well and deliver them stably; on the other hand, keep up with customers' next-generation connectivity solutions. Demand growth can bring orders, but after product forms change, whether past advantages can continue still has to be answered by new products and manufacturing capabilities.

This is also why this transaction is worth watching. It is not only an overseas company's expansion plan, but also shows us that competition in AI infrastructure is penetrating into more specific links. Chip performance is conspicuous, but connectivity, packaging and testing equally affect whether a system can operate efficiently. Chinese companies are already participating, and what they need to strive for next is to continue to hold a position in next-generation products.

Ultimately, what customers want is a system that can complete work efficiently and stably. Whoever can help it wait less and complete more tasks will have the opportunity to obtain sustained orders—whether what it provides is that expensive chip or the inconspicuous connections between chips.

This article comes from the WeChat public account "Tech Vortex", author: Wang Qinzhou, published with authorization from 36Kr.