Many people still look atAIand only stare at NVIDIA, as if the more chips there are, the stronger AI must be. After all, NVIDIA is still the core engine in the AI supply chain. The problem is that by today, AI data centers are no longer something you can solve just by buying a few GPUs. What will really separate winners from losers in the future may not be how powerful a single GPU is, but whether tens of thousands of GPUs can work together like a disciplined army.
You can think of GPUs as super-factories running at full productivity—each factory is expensive and highly capable. But if the highways between them are as narrow as country lanes, all the trucks pile up, warehouse dispatch gets chaotic, and these super-factories, no matter how powerful, can only wait for data and stare at each other. On paper, the compute looks luxurious; in reality, throughput is dragged down by transmission bottlenecks.
So today we're not talking about whether AI still lacks GPUs. We're talking about a more easily overlooked question for ordinary investors: once there are more and more GPUs, who solves the data-transmission bottleneck, and who makes these expensive chips truly connect into an efficient AI factory?
The key terms to remember are photonic chips, silicon photonics, optical modules, high-speed interconnects, network switching chips, DSP, and CPO, or co-packaged optics. Sounds like an engineering meeting, right? Every word seems familiar, but once they sit together you start pretending you don't know them. In fact, these are just the highways, bridges, tunnels, and traffic-control systems inside AI data centers.
In the first phase of the AI rally, money frantically chased GPUs because compute was the scarce resource then. Whoever supplied the strongest compute was king, so NVIDIA became the absolute core. In the second phase, the market began to realize that GPUs alone are not enough: power has to be sufficient, cooling has to hold up, servers have to be delivered, and data centers have to be built. So power, liquid cooling, servers, and data-center infrastructure all started getting attention.
So what is the third phase? If AI clusters keep expanding, the next wave of money will likely focus on high-speed interconnects. Because large-model training and inference are not a single GPU sitting alone in a server room fighting by itself; they are thousands of GPUs working together. Model parameters must be synchronized, training data must be transferred, inference requests must be distributed, and results must be sent back. The more GPUs you have, the more data they need to exchange with each other.
At that point, the efficiency of the AI system depends not only on the performance of one GPU, but also on the data-transfer speed, latency, bandwidth, and power consumption between GPUs. Put simply: compute is the engine, the network is the highway, and photonic chips and optical modules are the key bridges and tunnels inside that highway system. Without high-speed interconnects, no matter how many GPUs you have, you may still end up with a pile of expensive equipment stuck in the server room.
Why have photonic chips and optical communications suddenly been mentioned over and over by the market? Because traditional data centers also need networks, but AI data centers have completely different requirements. Old data centers handled webpages, video, file storage, and ordinary cloud computing; many tasks could be processed in a distributed way. The pressure was big, but it was still like normal city commuting.AItraining and large-model inference are different—they're more like suddenly hitting the Spring Festival travel rush, with every car trying to get onto the highway at once and every bit of data needing rapid synchronization. If one intersection gets blocked, the whole area slows down.
Copper wires, electrical signals, and traditional connections all run into bottlenecks in bandwidth, distance, power consumption, and heat dissipation. Optical communications have obvious advantages: faster speed, larger bandwidth, longer distance, and relatively lower energy use.AI data centers get larger and larger, and traditional electrical interconnects become more strained, so optical interconnects become more important.
That's why the market keeps talking about 800G, 1.6T, silicon photonics, and CPO. Not because these words sound advanced and are good for storytelling, but because AI data-center expansion really does require higher bandwidth, lower power, and lower latency interconnect solutions.
So how should you view this supply chain? You need to stay calm here—don't see the words "photonic chip" and just jump in. The biggest mistake in investing is not that you don't understand anything, but that you understand one word and think you understand the whole industry. That's when people get excited and buy in, and then when the market pulls back, their account gives them a risk lesson first.
The first layer is silicon photonics and optoelectronic integration. Simply put, that means combining optical functions with semiconductor processes to make optical communications more suitable for mass production and more suited to data-center scenarios. Broadcom, Marvell, Intel, and some photonic integration platform companies are all positioned in this direction.
The second layer is optical modules. This is the easiest to understand: servers, switches, and GPU clusters need optical modules to transfer data at high speed. 800G and 1.6T are the key upgrade directions. You can think of it as moving from a two-lane road to an eight-lane road, and then continuing to widen it in the future.
The third layer is network-switching chips and DSPs. Optical modules don't work alone; they need to cooperate with switching chips and signal-processing chips. Broadcom and Marvell are critical here, because they don't just sell a single component—they occupy key positions in the AI data-center networking upgrade.
The fourth layer is networking equipment and switches. AI clusters place very high demands on Ethernet and high-speed network architecture, so companies like Arista and Cisco will also be repriced by the market under the AI network-upgrade logic.
The fifth layer is CPO, or co-packaged optics. This sounds scary, but it's not hard to understand: you put the optical components closer to the switching chip, or even package them together. Why do that? Because the farther electrical signals travel, the greater the loss, the higher the power consumption, and the more annoying the heat dissipation. Moving optical components next to the chip is essentially about shortening transmission distance, lowering power consumption, and increasing bandwidth.
So this is not a single company eating the whole pie; it's a full AI data-center networking upgrade supply chain.
At the company level, Broadcom is more like a core asset. Its advantages are switching chips, custom chips, network architecture, and silicon-photonics ecosystems. Broadcom is not simply riding the AI theme; it's positioned at a key node in the AI networking upgrade. As long as AI clusters keep getting bigger, the market will continue to care about network and custom-chip demand. But Broadcom's problem is also clear: it's no longer a neglected asset, and its valuation already contains a lot of AI expectations, so you can't blindly chase it higher.
Marvell has more torque.It has positions in custom chips, DSP, optical interconnects, and data-center networking. If cloud providers and AI customers keep increasing demand for customized interconnects, Marvell's earnings leverage could become more obvious. But leveraged assets have one feature: they look great on the way up and are not gentle on the way down. They're suitable for watching earnings, orders, and guidance—not for chasing just because of emotion.
Coherent and Lumentum are more like hardware suppliers in the optical-device and optical-module chain. If 800G and 1.6T demand keeps exploding, they may benefit. But these companies have had cyclical characteristics in the past, and price competition plus margin volatility are key things to watch. Demand coming doesn't mean profits automatically come; more orders don't mean the stock price will hold up. Don't get confused here—we're investing, not watching a show and shouting slogans.
Arista is an important player in AI networking equipment. It may not always be classified as a photonic-chip company, but as AI data-center networking expands and the requirements for switches and network architecture rise, Arista naturally gets pulled into the AI network-upgrade logic. Its certainty is relatively stronger, but its valuation won't be cheap.
NVIDIA also can't be ignored. Many people think NVIDIA is just a GPU company, but NVIDIA has long been deeply positioned in AI network architecture.GPUs get stronger, and the network becomes more important. NVIDIA's push to expand AI clusters itself drives demand for high-speed interconnects. So NVIDIA is not only the main GPU theme, but also an important barometer for the whole AI interconnect space.
So why do I think this sector has investment value? There are five core reasons.
First, the larger the AI model, the higher the data-transfer demand. Small models may run on just a few servers, but large models depend on massive GPU clusters. The larger the cluster, the more critical interconnect becomes.
Second, the inference era will push up data traffic. Training is like building a skyscraper—centralized and project-based; inference is like city operations, with huge numbers of requests coming in every day. As AI moves into office work, search, advertising, customer service, coding, video, and robotics, data-center network pressure will shift from occasional spikes to a long-term high-load state.
Third, power consumption will become a hard constraint.AI data centers are already power-hungry beasts; GPUs consume power, cooling consumes power, and network transmission also consumes power. If the transmission solution is too power-hungry, the overall cost becomes absurd. The low-power advantage of optical communications will become more and more valuable in high-density data centers.
Fourth, the technology-upgrade cycle is already underway. From 400G to 800G, from 800G to 1.6T, and then to the longer-term CPO, this is not a short-term theme but a continuing upgrade cycle. Every bandwidth upgrade may drive equipment replacement, module upgrades, chip iteration, and a re-rating of the supply chain.
Fifth, capital needs to find a new direction beyond GPUs. NVIDIA is still the core, of course, but the market won't stare at one direction forever. As the GPU main theme rises enough, money will naturally look for the next bottleneck in the AI chain. Whoever solves the bottleneck may be repriced.
Still, don't get overexcited when you see this. Let me remind you: liking the direction doesn't mean you should rush in right away. The worst thing in investing is seeing the long-term logic correctly but buying at the top with short-term emotion. Others make money from the industry trend; you end up paying tuition to the market, and that tuition isn't cheap.
There are three things to watch closely in this sector.
First, look at the share of AI-related revenue. Not every optical-communications company can enjoy the AI dividend. If the company's main revenue is still traditional telecom, or if AI data centers make up only a tiny share, then you can't just slap an AI valuation on it. What really matters is the share of data-center revenue, AI customer orders, and the ramp-up of 800G and 1.6T products.
Second, look at gross margin. Optical modules and optical components have often been treated as cyclical manufacturing businesses because competition is fierce and pricing pressure is big. Revenue growth does not equal profit growth. Only when the mix shifts toward higher-end products, pricing power strengthens, and gross margins improve does real valuation re-rating make sense.
Third, look at the technology path.800G is the current focus, 1.6T is the next stage, and CPO is the longer-term direction. If a company lacks the ability to work with next-generation high-speed interconnect and silicon-photonics platforms and is only rebounding on old products, then it may just be a sentiment bounce, not a long-term re-rating.
For the next move, I lean toward three possibilities.
First, photonic chips and high-speed interconnect become the new AI main theme. The prerequisite is that NVIDIA, cloud providers, and server companies keep validating demand for AI data centers, while Broadcom, Marvell, Coherent, Lumentum, and Arista show order improvement and guidance upgrades in earnings. Then money may spread from GPUs to the interconnect sector.
Second, a structural rally. Not the whole industry rises together, but the companies with the most certain orders, the highest AI mix, and the clearest technology path lead. Broadcom, Marvell, and Arista are more core; Coherent and Lumentum are more leveraged. Core assets are steadier, leveraged assets are more exciting, but they can also feel like a roller coaster. If you're risk-averse, don't pretend otherwise—the account will tell the truth for you.
Third, a short-term catch-up rally. If AI data-center capex doesn't keep being revised upward, or if optical-module price competition intensifies, or if market risk appetite weakens overall, then this line may just run on sentiment for a while and then re-diverge.
Risks also need to be on the table.
First, technology-path risk. Silicon photonics and CPO have huge imagination, but commercialization may lag behind stock-price expectations. Second, price competition risk. Good demand does not mean good profits; even if orders arrive, price cuts can eat the margin. Third, customer concentration risk.AI data-center customers are concentrated among a handful of cloud providers and tech giants, and once their capex pace changes, orders will swing visibly. Fourth, valuation risk. Once AI-related stocks have risen too much, market expectations become very demanding. Even a good earnings report may not be exciting enough and the stock can still fall. Fifth, macro risk. If U.S. Treasury yields rise or worries about tighter liquidity reappear, high-valuation tech stocks will come under pressure, and photonics won't be immune.
Ordinary investors should keep an eye on five signals next.
First, watch Broadcom and Marvell's guidance. If they keep giving strong guidance in AI networking, custom chips, and optical interconnects, then this main theme is being confirmed by earnings.
Second, watch Coherent and Lumentum. If they don't just spike for a day but continue to trade strongly on volume, and if earnings show data-center revenue and gross margin improvement, then the market may be repricing optical-device demand.
Third, watch whether Arista keeps outperforming the broader market. If AI networking equipment companies stay strong, it means money isn't just trading a concept, but is buying the data-center network upgrade.
Fourth, watch the relative strength between NVIDIA and photonics names. If NVIDIA consolidates at high levels while optical interconnect names get stronger, that suggests money may be spreading from GPUs into interconnects.
Fifth, watch the internal structure of the Nasdaq and the semiconductor sector. If the broader market is choppy but optical communications, networking, and silicon-photonics names move against the trend, that means the market isn't leaving AI—it's searching for the next bottleneck inside AI.
In summary, AI development does not just push GPUs; it pushes the entire data-center infrastructure stack. Power solves whether there is enough electricity, cooling solves whether heat can be removed, servers solve whether GPUs can run stably, and photonic chips plus high-speed interconnect solve whether those GPUs can truly work together efficiently.
AI's first phase buys compute, the second phase buys data centers, and the third phase will likely focus on high-speed interconnect. Photonic chips are not just chasing the AI hot topic; they are the inevitable bottleneck as AI clusters keep expanding. The stronger the compute, the more critical interconnect becomes; the more GPUs you have, the more important optical communications get; the larger the data center, the greater the imagination for silicon photonics and CPO.
But the last word is still the same: we care about industry trends, not emotional charging. Mature investment decisions are not about seeing a new concept and going all in; they're about looking at the industry logic first, then the earnings data, then the valuation, and only then the entry point. Getting the direction right is only the first step; timing it right is the real skill.
If later you see NVIDIA consolidating at high levels, but Broadcom, Marvell, Coherent, Lumentum, and Arista—companies tied to AI networking and optical communications—begin to rise together, that may mean money isn't leaving AI but is looking for the next bottleneck in the AI chain. Those who missed GPUs may not keep chasing the already-expensive NVIDIA; instead, they may shift to AI interconnect, optical modules, silicon photonics, and CPO to look for catch-up opportunities.
Which company do you like best? Broadcom, Marvell, Coherent, Lumentum, Arista, or NVIDIA? You can drop the ticker directly in the comments. If this episode helped you understand the underlying logic of AI photonic chips and high-speed interconnect, remember to like, save, and subscribe. This is Frost Financial, and next time we'll continue to use easy plain language to break down complex U.S. stock market moves into investment logic ordinary people can understand.
