🔥 The one main theme that has made me increasingly certain recently is:

The next bottleneck for AI isn't just GPUs, it's optical communication.

Many people are still focused on NVDA, AVGO, MRVL, but if you break down the industry chain, you'll find that the real big opportunities often lie in the deeper layers of the supply chain.

Silicon Photonics (SiPh) and Co-Packaged Optics (CPO) essentially solve one problem:

As computing power gets stronger, how does data flow at high speeds between chips, servers, racks, and data centers?

In the past, the core bottleneck for AI was computing power.

Now, it's starting to become bandwidth.

In the future, it will definitely become power consumption.

No matter how powerful the GPU is, if data can't be transmitted, cluster efficiency won't improve.

HBM solves "whether the data can be fed fast enough."

CPO and optical interconnects solve "whether the data can move fast enough."

This is why I'm paying more and more attention to this chain:

HBM → Optical Modules → Silicon Photonics → CPO → Optical Interconnects

This is the inevitable path for AI infrastructure upgrades.

The most certain direction isn't necessarily the most exciting story, but who holds the key to the physical bottleneck.

For example:

Light source end: watch $SIVE, $Lumentum(LITE.US), $Coherent Corp.(COHR.US), $Broadcom(AVGO.US), $Applied Optoelectronics(AAOI.US).

Silicon photonics foundry: watch $Taiwan Semiconductor(TSM.US), $GlobalFoundries(GFS.US), $United Microelectronics(UMC.US), $Intel(INTC.US).

Digital coherent optics: watch $Nokia Oyj(NOK.US), $Ciena(CIEN.US), $Cisco(CSCO.US), $Coherent Corp.(COHR.US).

Optical interposer: watch $POET Tech(POET.US).

Analog/mixed-signal ICs: watch $Marvell Tech(MRVL.US), $Macom Tech(MTSI.US), $Semtech(SMTC.US), $Maxlinear(MXL.US).

Connectors and fiber optics: watch $Corning(GLW.US), $Amphenol(APH.US), $TE Connectivity(TEL.US).

Test and measurement: watch $Formfactor(FORM.US), $Keysight Tech(KEYS.US), $Viavi Solutions(VIAV.US), $Aehr Test(AEHR.US).

Advanced packaging: watch $BESI, $Onto Innovation(ONTO.US), $Camtek(CAMT.US).

This isn't just hype.

This is the engineering bottleneck that will inevitably be hit as AI computing power continues to expand.

My own understanding is:

Stage one, the market hyped GPUs.

Stage two, the market hyped HBM.

Stage three, the market will definitely hype optical interconnects.

Because as GPUs proliferate, servers become denser, and data centers grow larger, the real limiting factor for efficiency will shift from "how fast a single chip computes" to "how fast the entire system is connected, how low its power consumption is, and whether its cooling can handle it."

So, when I look at this line now, I divide it into two categories:

The ones with stronger certainty are $Marvell Tech(MRVL.US), $Nokia Oyj(NOK.US), $Coherent Corp.(COHR.US), $Corning(GLW.US), $Amphenol(APH.US).

The ones with higher potential returns but greater volatility are $SIVE, $POET Tech(POET.US), $Lightwave Logic(LWLG.US).

The former are core players in the industry chain.

The latter are small-cap stocks with elasticity and technological imagination.

The most interesting thing about this line is:

It's not rising purely on stories, but is driven by the actual expansion of AI infrastructure.

If AI continues to expand in the future, data centers can't just stack GPUs.

They must upgrade storage, optical communication, connectors, packaging, testing, and power management.

That's why I always say:

Left foot storage, right foot optical communication.

Storage determines whether AI can be fed.

Optical communication determines whether AI can run far.

Truly big market trends often don't start when everyone sees them.

But when the bottleneck has just been exposed and the industry chain hasn't been fully priced in yet.

Right now, silicon photonics and CPO might be in that position.