Another long piece on optical modules and CPO crossed my screen this week. The core claim was simple: as AI scales, the bottleneck stops being just GPUs and increasingly shifts toward interconnect.
I agree with that.
Where I disagree is how quickly that statement gets flattened into an investment conclusion. Once the conversation moves from technology to stocks, three very different things are often bundled together:
- revenue already showing up in orders
- technology paths that may validate over the next two to three years
- a longer-duration endgame narrative
That is how an entire sector gets sold as one clean opportunity set, as if anything touching optical modules, CPO, or silicon photonics automatically belongs in the same bucket.
In markets, it does not work that way.
The right way to frame this chain is around three questions:
- What is the actual relationship between optical modules, CPO, and silicon photonics?
- Which companies are closer to real economic exposure, and which are mostly narrative proxies?
- Why is the better framework not “believe in the theme,” but “find the bottleneck”?
Why this has become a core AI infrastructure trade
For the past two years, the default market narrative has been straightforward:
- AI demand exploded
- GPUs became scarce
- compute was everything
That logic was directionally right, but incomplete.
Once training and inference clusters become large enough, the question changes from how many GPUs you own to how efficiently those GPUs talk to one another.
Think of a large AI cluster as a city:
- GPUs are factories
- HBM is warehousing
- CPUs are dispatch centers
- networking is the road and rail system
If you keep adding factories and storage, but the transport grid becomes congested, power-hungry, and slow, the system's marginal efficiency starts to degrade very quickly.
That is why the market is now repricing “connectivity” as infrastructure rather than as an accessory. Interconnect is no longer just attached to the compute story. It is part of the compute system.
That is also why investors keep circling the same set of themes:
- optical modules
- optical components
- lasers
- silicon photonics
- CPO
- and eventually broader optical networking architecture
This trade is not interesting because it is new. It is interesting because it sits closer to a real systems bottleneck.
Optical modules, CPO, and silicon photonics are not the same thing
These terms are constantly grouped together, but they sit at different layers of the stack.
1. Optical modules: the current cash-flow line
Optical modules are, at the simplest level, electro-optical converters.
Chips and boards operate in electrical signals. High-speed transmission over distance increasingly prefers optical signals. The module is the bridge between the two.
From an investment perspective, optical modules matter because they have three characteristics:
- the cleanest demand logic
- the clearest order visibility
- the fastest path to monetization
If AI data centers continue to scale and the industry keeps stepping through 400G -> 800G -> 1.6T, module vendors are generally first in line to see real revenue.
That is why leadership in the group tends to show up first in optical modules, not in the furthest-out end-state narrative.
2. CPO: a medium-term architecture path, not an immediate earnings stream
CPO stands for Co-Packaged Optics.
The idea is to move optics closer to the ASIC or switch silicon, and in some cases package them together, instead of relying on fully pluggable modules at the board edge.
The attraction is obvious:
- lower power
- lower latency
- higher bandwidth density
- shorter high-loss electrical paths on the board
That makes CPO easy to romanticize. Investors hear “next-generation architecture” and instinctively jump to “next major equity winner.”
But a correct architecture path and a near-term earnings stream are not the same thing.
CPO still faces a set of practical constraints:
- packaging complexity
- yield management
- thermal control
- serviceability and replacement
Below that sit the usual industry frictions: standards, interoperability, and platform control.
So the useful question is not whether CPO is strategically right. It probably is. The useful question is who gets paid during the transition.
3. Silicon photonics: the technology substrate, not a free pass
Silicon photonics is closer to a technology substrate than a single product category.
It is about integrating more optical functionality at the chip, device, and engine level to support higher bandwidth, lower power, and denser system design.
From an industry perspective, it is compelling.
From an equity perspective, it is also the part of the story most likely to be over-discounted early.
Why?
- the technology is attractive
- the narrative is ambitious
- commercialization usually moves slower than the write-up
My view is simple: silicon photonics matters, but investors need to separate long-duration platform optionality from short-duration earnings certainty.
The right way to read the value chain
If I had to compress the chain into one sentence, it would be this:
Optical modules are the present cash-flow layer, CPO is the mid-cycle validation layer, and silicon photonics is the long-term end-state layer.
Under that framing, the stack looks roughly like this:
| Layer | What is being sold | What matters now | Who tends to benefit |
|---|---|---|---|
| Platform definition | GPUs, switch ASICs, network architecture | roadmap control and standards power | NVDA AVGO MRVL |
| Optical modules | 400G / 800G / 1.6T modules | orders, speed upgrades, customer ramps | module leaders |
| Optical components | coupling, connectors, AWG, passive devices | volume growth and content per system | “picks-and-shovels” suppliers |
| Active devices | lasers, detectors, EML | process know-how and scarce supply | upstream core device suppliers |
| Silicon photonics / optical engines | PICs, modulation, waveguides, coupling | technical validation and platform fit | genuine SiPh participants |
| CPO / NPO packaging | near-package or co-packaged optics | yield, thermal performance, scaled delivery | transition beneficiaries |
| Advanced packaging | heterogeneous integration, SiPh manufacturing | production cadence and ecosystem alignment | foundry and advanced packaging ecosystems |
What looks like one value chain is really three different pools of money:
- order money
- validation money
- end-state money
A lot of investors lose money not because the theme is wrong, but because they pay for the second and third bucket as if they were buying the first.
Why “find the bottleneck” is better than “chase the hot concept”
One framework I find consistently useful is this:
Do not start by asking who the leader is. Start by asking where the bottleneck is.
That lens is especially useful for optical modules, CPO, and silicon photonics.
1. Start with demand, not price action
Is AI data-center demand still expanding? Most likely, yes.
But that is not the highest-value question.
The better question is which part of the system gets pushed to its limit first. Not every layer benefits at the same time, and the market usually re-rates the layer where the constraint bites first.
2. Excess returns usually come from scarce supply
Demand growth by itself is not enough. Scarcity is what creates economics.
If everyone can make a product, rising demand often becomes a pricing fight.
If only a small number of companies can deliver it, customer qualification takes time, and yield ramps are slow, that layer has a better chance of earning excess profits.
In this chain, the areas worth watching most closely are usually:
- core supply in high-speed modules
- upstream active devices such as lasers and detectors
- silicon photonics-enabled optical engines
- high-end packaging and component chains tied to CPO or NPO
3. A plausible story is not the same as a monetizable one
This is where A-shares and Hong Kong names can get tricky.
A company can sit close enough to the narrative to be marketed into the theme while still being far from actual monetization. The logic may sound coherent. The industrial adjacency may be real. But revenue and margin realization can still be distant.
Those names can trade very hard in a hot tape. That does not make them the highest-conviction beneficiaries.
In A-shares and Hong Kong, who looks closer to real exposure?
This is a framework discussion, not a claim of precision on every single name.
Category one: closer to real beneficiaries
1. Zhongji Innolight
This is the clearest example of a current-order beneficiary.
The core thesis is not the abstract CPO endgame. It is direct exposure to high-speed optical module demand from AI data-center buildout.
2. Eoptolink
Broadly the same logic applies here. If the trade is about speed upgrades and data-center expansion, direct module exposure is usually more grounded than concept-heavy proxies.
3. Tianfu Communication
I tend to think of this as a stronger picks-and-shovels business. If the optical chain keeps scaling, it can benefit from both volume growth and higher value per system.
4. Accelink
This is closer to an established platform participant than a late-stage thematic name. The upside may not always be the most explosive, but the industrial position is more tangible than many newer concept stories.
5. HGTECH
This belongs more naturally in the NPO / silicon photonics / transition-beneficiary bucket. If domestic routing leans toward near-package and gradual upgrade paths first, it has a more concrete industrial handle than a large group of pure slide-deck stories.
6. Shijia Photons
This sits closer to upstream optical chips and component leverage. It is less near-term earnings-like than the module leaders, but still more real than many names that simply borrow the narrative.
7. Hong Kong names tied to optical devices or silicon photonics
If the goal is to find businesses closer to the technology evolution itself, that subset is generally more interesting than broad AI-communications proxies. The trade-off is that liquidity, investor base, and valuation behavior in Hong Kong are different from A-shares.
Category two: closer to narrative mapping
By “narrative mapping,” I mean names that can be connected to the story conceptually but are still far from being core economic beneficiaries.
Typical features include:
- no real 800G or 1.6T product traction
- no top-tier customer validation
- no scarce advantage in lasers, PICs, packaging, or yield-critical processes
- legacy revenue still dominating, with AI optics added mainly at the narrative layer
My view on these names is straightforward: they can be tradable as sentiment, but they should not be confused with core industrial assets.
The three most common mistakes in this trade
Mistake one: buying the end-state as if it were current earnings
The market hears CPO or silicon photonics and immediately extrapolates the future into the present.
But between architecture and earnings there is still a long list of gating variables:
- process maturity
- yield
- customer qualification
- production scaling
- standards alignment
Any one of those can slow the path from thesis to P&L.
Mistake two: assuming the whole sector benefits equally
Yes, the broader optical-AI chain benefits.
No, the benefit is not evenly distributed.
Some names earn on current orders, some trade on validation, and some mostly trade on sentiment. Those buckets should not be valued the same way.
Mistake three: confusing storytelling strength with bottleneck position
The companies talking most aggressively about CPO or silicon photonics are not necessarily the ones controlling the hardest part of the stack.
What matters is whether a company sells the piece that is most scarce, hardest to replace, and first to be amplified by the system bottleneck.
The practical conclusion
If I break the trade down by time horizon, this is how I would frame it:
Short term
Focus on optical modules + optical components + lasers + packaging-adjacent beneficiaries.
That is the part of the chain closest to order conversion.
Medium term
Focus on NPO / CPO validation.
Not who talks the loudest, but who actually enters platform ecosystems, wins customer qualification, and proves yield plus delivery.
Long term
Focus on silicon photonics and higher-density optical networking architectures.
The opportunity is large, but many names here should be treated as long-duration optionality rather than current certainty.
If I had to reduce the whole discussion to one sentence, it would be this:
The real trade is not “believe in the light.” It is to follow the bottleneck and ask who gets paid now, who gets paid on validation, and who is still selling a distant story.
The theme deserves attention.
But in markets, getting the direction right is only the start. The harder part is getting the timing, the security selection, and the valuation framework right.