@Gaetano2026 ↔ @johhnyWalkerAZ conversation

We are so early on Photonics.
$LITE $POET $AAOI $COHR $MRVL

There is so much to learn about Photonics/Optics
I learn more and more everyday

A major goal of my page is to be a place where people can come and learn both about the tech and the companies poised to capture market share

So I want to open up a post here to just have a discussion/ Q&A

Ask me ANYTHING here
From the fundamentals to the more technical

I will do my best to answer all the questions and pull references when needed. And if I don't feel confident in my response, I will state that and do more research

Please just be specific if you can!

Drop any questions you have!

Hey Johhny,

There is a lot to unpack here. I’ll start with a quick answer and then dive into more detail.

In short, optics is already widely used at longer distances and it is progressively targeting the shorter distances that copper is still highly effective in and sought after.

So let’s start by explaining what copper is and what optics is and the strengths and drawbacks of each and then where I see the industry moving towards.

Copper is an electrical link. It’s really inexpensive and it doesn’t require you to convert the signal into light, which is a big reason it has worked so well for so long at short distances. Also important, copper comes in two modes in practice. There is passive copper (DAC), which has no active electronics in the ends, so the cable itself adds essentially no incremental power draw. And there is active copper, where you still use a copper cable but you add signal-conditioning electronics in the ends so it can run farther at high speeds. I’m pointing this out now because it matters a lot when we get to scale out in a few paragraphs.

Optics is capable of handling really high speeds with low signal loss because it uses light instead of electricity. It is working really well for longer distances like connecting 2 data centers or connecting different racks that are far apart. While copper signals start to fall apart once you push high speeds beyond short reach, fiber can carry that same data for kilometers depending on the optics and link type, and it does it without fighting the same electrical channel problems.

A few key terms that we need to talk about before discussing more specifics on what the future looks like for copper and optics involve how we scale these systems.

'Scale up' refers to the connecting of many GPUs within a single rack so they act like one giant computer. In the latest NVIDIA Blackwell NVL72 racks, they use on the order of thousands of copper links (Jensen has described it as roughly 5,000 cables totaling about 2 miles of copper) to link 72 GPUs. This is a very short distance, typically under 1 meter, and this is where copper dominates. The reason is density and power. If you tried to make every one of those ultra-short links optical today, you would stack a lot of conversion power and heat inside the rack, and the rack-level penalty can be on the order of tens of kw.

'Scale out' refers to the connecting of many GPU racks to each other across a data center floor. This could be anywhere from 2m to 100m or so depending on the layout. This is the main territory where optics is starting to push copper out, but it’s not a simple on/off switch. At 800G and as we move toward 1.6T, passive copper tends to be limited to the shortest reaches because the signal margin collapses quickly with length. That is where you hear the 1 to 2 meter type ranges. But active copper is a real middle ground. That is where AECs and retimers show up, and they can keep copper viable for several meters in the exact same scale out domain before optics becomes the obvious answer.

'Scale across' refers to the connecting of GPUs across different data center buildings. This could be upwards of a few kilometers. At this distance, copper physically cannot carry the signal, so optics is the way to do it.

Also, copper has tons of drawbacks at high speeds like loss, reflections, and crosstalk (where electrical signals interfere with each other). Even when the cable is passive, the system still ends up paying for a harder electrical channel with more equalization and tighter margins as speeds climb. To extend copper’s life and distance, the industry has created innovations like AECs (Active Electrical Cables) and retimers.

AECs are basically copper cables with tiny signal-conditioning chips in the ends. These chips clean up the signal, allowing copper to reach several meters at speeds where passive copper would struggle, and in some designs you can get up into the mid to high single digit meter range. This allows engineers to keep using cheaper copper for a little while longer even as speeds double, but now you are paying more than passive copper in both cost and power to buy that extra reach. There is also a move toward LPO (Linear Pluggable Optics), which is a way to reduce the power of fiber optics by simplifying the electronics in the module. The tradeoff is you get tighter link requirements and less margin, but the goal is for lower-power optics that can move inward faster.

The whole idea a lot of the industry references is to use copper when you can and optics when you must. Optics is really expensive relative to copper (for now. Alpha in identifying the companies that can change this) and it comes with drawbacks like power and heat per endpoint, reliability and qualification requirements, and serviceability. (Again, there is a lot of opportunity in identifying the companies that can make optics more feasible and whom target these drawbacks). A passive DAC is still the cheapest, simplest link you can deploy when it works. But once you need more reach at high speed, you either pay up for active copper as a bridge, or you move to optics and stop fighting the electrical channel.

However, at certain speeds and distances, copper just can’t hang. That’s what we are seeing with optics dominating scale across and the longer distances in scale out. As time progresses I see optics continuing to chip away at copper’s market share to the point where copper will be used in only really specific applications in the data center that just don’t need optics, mostly the ultra-short, ultra-dense links where power and simplicity matter more than anything. There is always a disruptor on the horizon though.

This will take a long time but one example that you have heard of with this is CPO (Co-Packaged Optics). This is the mid term end game direction where we move the optical engine much closer to the switch or compute package so you are not pushing fragile electrical signals across longer copper paths. The early real deployments are more likely to show up on switching platforms first, and then the industry works toward moving optical I/O closer to compute over time. I expect this to start hitting the market in a meaningful way around 2027, but it will be an adoption curve in steps, not an overnight replacement of copper everywhere.

I hope this answers your question! There is a lot more to discuss here but this lays a good foundation without getting too much in the weeds