3 AI Networking Stocks Quietly Dominating Their Niche in August - 24/7 Wall St.

Skyler: Welcome to Line Rate — the show about datacenter fabric and A I high-speed interconnect. I'm Skyler Monroe, and if it moves bits between accelerators, racks, or datacenters, we're talking about it here. Big thanks to our sponsors Ada, who helps clients integrate A I into their workflows, and Zen Semiconductor, an umbrella company with a portfolio of companies across the silicon industry. Today we've got a lot to unpack — Google just dropped something called Virgo that could flip how we think about scale-out fabric, the co-packaged optics world is getting very busy with Molex, Teramount, Ciena, and some genuinely wild photonics research, and we'll touch on which A I networking stocks are quietly pulling ahead. Let's get into it. Skyler: Alright, first up — a story from 24/7 Wall Street about three A I networking stocks that are, quote, quietly dominating their niche. Now I know, I know — stock picks on a technical show, bear with me, because the reason these names keep surfacing actually tells you something real about where the infrastructure money is flowing. The piece calls out companies in the ethernet switching, optical transceiver, and high-speed cable assembly spaces as the ones pulling ahead without a lot of fanfare. And honestly, that tracks. The loudest names in A I infrastructure right now tend to be the GPU vendors and the hyperscalers themselves, but the boring-sounding middle layer — the switches, the optics, the cables — that's where the actual capacity constraints live. If you're building a 100,000-GPU cluster, you need switching bandwidth that scales with it, and right now the demand curve for that gear is essentially vertical. The story doesn't get super deep on the technical side, which is fine, it's a finance publication, but the implication is clear: the companies that own the plumbing between accelerators are in a really strong position. What I find interesting from an engineering standpoint is that these niches are converging. Ethernet switch vendors are racing to support 800G and now 1.6T per port. Optical transceiver vendors are dealing with power density nightmares as pluggable optics push into the 40-plus watt range per module. And cable assembly shops are trying to keep up with DAC and active copper specs that change every 18 months. The common thread is that all of them are being pulled forward by A I training cluster builds, and the hyperscalers are signing long-term supply agreements to lock in capacity. So when a financial outlet says these stocks are "quietly dominating," what they're really saying is the people who actually build datacenters have already made their bets, and the contracts are in place. That's not nothing. It's a signal that the infrastructure layer is maturing faster than the public narrative gives it credit for. Skyler: Okay, now the one I've been most excited to dig into this week — Google unveiled something called Virgo, and it's a genuinely interesting architectural move. The headline is that Virgo is a scale-out fabric that ditches the traditional three-layer network to unify A I accelerators. So let's break that down because there's a lot packed in there. Traditional datacenter networks for A I training have followed a pretty standard spine-leaf-or-three-tier model — you've got your top-of-rack switches, your aggregation layer, your spine, and the A I accelerators hang off the edges. It works, it's proven, but it introduces latency and bandwidth bottlenecks at every layer boundary, and when you're doing all-reduce across tens of thousands of GPUs, those boundaries really start to hurt. What Google is saying with Virgo is — what if we just didn't have those layers? The idea is to build a unified, flat fabric that directly interconnects accelerators without the hop-by-hop inefficiency of a tiered topology. Now this is not a brand new concept in principle — dragonfly and flattened butterfly topologies have been discussed in academic literature for years, and we've seen some of this thinking in HPCs. But Google doing it at hyperscaler scale, with their own custom silicon and TPUs in the mix, is a different thing entirely. The name Virgo is interesting too — Google has a pattern of naming their networking projects after constellations and celestial references, and Virgo fits the family. What makes this technically significant is the claim that it unifies accelerators — meaning TPUs and likely other compute nodes — under a single fabric abstraction. That's not just a topology change, that's a control plane and routing challenge of serious magnitude. How do you do collective communication primitives efficiently on a flat fabric at that scale? How do you handle failures without a hierarchical fallback? These are hard problems. Google has the engineering depth to solve them, but it also means this isn't something you can go buy off the shelf — Virgo is a vertically integrated Google thing, at least for now. The broader implication for the industry is that if Virgo works well, it puts pressure on everyone selling tiered ethernet fabrics for A I. It's another data point in the "is ethernet good enough for tight collectives" debate, alongside Ultra Ethernet Consortium's work on U E C. Google essentially saying "we built our own thing" is a signal that the standards bodies may not be moving fast enough for the hyperscalers' actual needs. Skyler: Moving on to co-packaged optics, which has been quietly having a very loud month. Let's start with the Molex acquisition of Teramount, because this one has some real strategic logic behind it. Molex is a well-established interconnect company — connectors, cables, that whole world — and Teramount is an Israeli photonics startup that has been working on silicon photonics coupling technology. Specifically, Teramount has a really interesting approach to the problem of getting light efficiently from a laser or fiber into a silicon photonic chip. That coupling problem — sometimes called the fiber-to-chip interface — is one of the gnarliest manufacturing challenges in the whole co-packaged optics stack. The losses at that interface can be significant, and if you're trying to co-package optics right next to a switch ASIC or an A I accelerator, you need that interface to be reliable, low-loss, and manufacturable at scale. Teramount's approach uses what they call a photonic tunnel, which essentially creates a very efficient evanescent coupling path that's more tolerant of the alignment errors you get in high-volume manufacturing. That's the killer feature — not just raw efficiency, but yield-friendly efficiency. For Molex, this acquisition makes a lot of sense. They're already in the cable and connector business that feeds into co-packaged optics deployments, and owning the fiber-to-chip coupling IP means they can play further up the stack. It also tells you something about where Molex thinks the volume is going — they're betting that co-packaged optics goes from research curiosity to mass deployment fast enough to justify the acquisition premium. And looking at the roadmaps from the major switch vendors, that bet looks pretty well-founded. Skyler: The EE Times piece this week takes a broader view and it's worth spending a few minutes on because it zooms out from the component level to ask — where is photonics actually going inside the package? The story traces a trajectory from pluggable optics, which most people in this audience know well, through co-packaged optics, and then toward something even more integrated — nanolasers and on-chip photonic integration at a level that starts to blur the line between the optical and electronic domains. The key insight in the piece is that every generation of this trend is about moving the optical function closer to the compute, and each move reduces the electrical loss you take getting the signal from the chip to the optical interface. Pluggable optics are sitting on the front panel of your switch, connected by a relatively long electrical trace. Co-packaged optics moves that into the same package as the switch ASIC, cutting that trace length dramatically. But nanolasers and on-chip photonics push the integration even further — we're talking about light sources that could potentially live on the same die or directly bonded to the compute die. Now, that last step is still very much research territory. The physics are hard, the thermal environment on a compute die is brutal for lasers, and the manufacturing integration is a genuinely unsolved problem at scale. But the direction is clear, and the reason it matters for A I fabric specifically is bandwidth density. The amount of data you need to move in and out of an A I accelerator is growing faster than the electrical signaling budget can handle. Optics is the long-term answer, and how close you can get the optics to the die determines how efficiently you can use the available bandwidth. So when you see a story about nanolasers, don't think of it as sci-fi — think of it as the five-to-ten year roadmap answer to a problem that's already showing up in today's cluster designs. Skyler: And rounding out the co-packaged optics coverage this week — Ciena introduced what they're calling a pluggable optical engine designed to help accelerate co-packaged optics adoption. And I want to spend a minute on this because the framing is a little counterintuitive. Ciena is a company known for long-haul and metro optical networking — coherent transmission, DWDM, that world. Co-packaged optics is typically a short-reach, intra-datacenter play. So what's Ciena doing here? The answer is that the "pluggable optical engine" concept they're proposing is essentially a bridge technology — it takes some of the optical integration work that co-packaged optics requires and packages it in a way that's more accessible to switch and accelerator designers who aren't ready to go full co-packaged. Think of it as a modular optical engine that plugs into the system but handles more of the optical complexity than a traditional transceiver, giving you some of the power and density benefits without requiring you to redesign your entire package. The strategic angle for Ciena is interesting — they're using their optical systems expertise to enter the datacenter fabric optics market, where the volume is enormous and growing fast. The co-packaged optics ecosystem is still relatively early, and there's a real interoperability question around how different vendors' components work together. By positioning a pluggable optical engine as an on-ramp, Ciena is essentially saying "you don't have to go all-in on co-packaged day one, here's a stepping stone." Whether that stepping stone becomes a durable product line or just a transitional technology that gets absorbed into full co-packaged deployments — that's the interesting question. But either way, it's another sign that the photonics integration wave is pulling in companies from adjacent markets who see the datacenter as the next big optical battlefield. Skyler: Alright, that's the show for this week. We covered Google's Virgo fabric and what flat-topology scale-out means for the A I networking industry, dug into the Molex-Teramount acquisition and why fiber-to-chip coupling is such a critical piece of the co-packaged optics puzzle, looked at the long arc from pluggable optics to nanolasers, and talked through Ciena's new pluggable optical engine as a bridge into the co-packaged world. Thanks again to our sponsors — Ada, who helps clients integrate A I into their workflows, and Zen Semiconductor, an umbrella company with a portfolio of companies across the silicon industry. If you found this useful, please subscribe wherever you get your podcasts, and share it with someone who cares about the infrastructure layer — there are more of us than LinkedIn would have you believe. Full show notes with links to all the stories we covered are available at the website. See you next week on Line Rate.