AI Data Center Connectivity Wars: Copper and Optics Not Substitutes, CPO Reshapes Industry Profit Landscape - finance.biggo.com

Skyler: Hey, welcome to Line Rate — the show tracking everything happening in datacenter fabric and A I high-speed interconnect. I'm Skyler Monroe, and if it moves bits between chips, racks, or datacenters, we're covering it. Quick thanks to our sponsors: Ada, who helps organizations actually integrate A I into their workflows in a way that sticks, and Zen Semiconductor, an umbrella company with a portfolio of businesses across the silicon industry. Really glad to have them both on board. Alright — today we've got a lot to get into. The copper versus optics debate is getting more nuanced, Molex just made a move to pick up a photonics startup, photonics is pushing deeper into the package than ever before, and Ethernet just body-slammed InfiniBand in the A I fabric market share numbers. Let's go. Skyler: Okay, first up — there's a piece making the rounds about the connectivity wars inside the A I datacenter, and it's pushing back on something I hear constantly, which is this framing of copper versus optics as if they're fighting over the same territory. The argument here, and I think it's largely correct, is that copper and optics aren't substitutes — they're complementary, and the real action is about where the boundary between them sits and who captures the margin at that boundary. Skyler: Here's the technical reality. Copper — whether we're talking passive DAC cables or active electrical cables — is still dominant within the rack and for very short reaches. Sub-two meters, you're almost always on copper because the power and cost math just works out. Optics historically took over somewhere in the one-to-three meter range depending on the generation and the signal rate, and that boundary has been creeping outward as AEC technology improves. Active electrical cables with retimers can now push further than people expected even a few years ago. Skyler: But here's where co-packaged optics — CPO — completely reshapes the conversation. CPO moves the optical engine from a pluggable module on the faceplate to right next to the switch ASIC on the package itself. And when you do that, you dramatically cut the electrical trace length that the high-speed signal has to travel before it hits photons. That's a huge deal because SerDes power scales badly with reach and data rate. You're basically trading a long lossy electrical path for a very short one, and then converting to light right there at the package. Skyler: The industry profit angle in this story is what I find really interesting. When CPO becomes the dominant architecture — and I do think it will for merchant silicon at scale — the margin picture shifts significantly. Right now the pluggable optics vendors, your Coherent, your II-VI, your Inphi-era players, they capture a ton of value per port. CPO threatens to pull some of that value back toward the ASIC vendor or toward whoever is integrating the optical engine at the package level. So this is not just a tech transition, it's a supply chain and margin war. And that's why you're seeing so many acquisitions in this space right now, which brings me perfectly to the next story. Skyler: So Molex — yes, the connector company, though they are way more than that at this point — has announced they're acquiring Teramount. And if you haven't heard of Teramount, they're an Israeli photonics startup with some genuinely interesting technology around fiber-to-chip coupling. That last piece, getting light efficiently from a fiber into a silicon photonics waveguide, is one of the gnarliest unsolved problems in the CPO stack, and Teramount has been working on it with what they call a photonic busbar approach. Skyler: The core challenge with fiber-to-chip coupling is alignment. Single-mode fiber has a mode field diameter of around eight to ten microns. Silicon photonics waveguides are much smaller than that — we're talking sub-micron in some dimensions. Getting light from one into the other without wasting a ton of it requires either very precise active alignment, which is expensive and slow, or some kind of passive alignment scheme that's manufacturable at scale. Teramount's approach uses a lens-based coupling structure that's supposed to enable passive alignment with reasonable insertion loss. If that works at volume, it's a big deal. Skyler: For Molex, this makes a lot of strategic sense. Molex already plays in the high-speed connector and cable space, they're inside a lot of datacenter deployments, and they've clearly been watching the CPO transition coming. Owning a fiber coupling technology means they can potentially be the company that connects the external fiber plant to the CPO module on the package. That's a critical interface point and it's one where nobody has really locked up the high-volume solution yet. This acquisition is Molex planting a flag and saying: we want to own that piece. Skyler: From an ecosystem standpoint, this is also a signal to watch. When a company like Molex — which is itself owned by Koch Industries, so there's significant capital behind them — starts acquiring photonics IP, it tells you the whole connector and cabling industry is taking CPO seriously as an existential challenge to the pluggable module business. The smart money is not fighting CPO, it's figuring out where you fit in the CPO world. And Teramount gives Molex a real technical answer to that question. Skyler: Alright, story three — EE Times ran a great piece looking at the broader arc from co-packaged optics toward what they're calling the next frontier: nanolasers. And I want to spend some time on this because I think people sometimes hear "nanolaser" and think it's science fiction. It's not. It's closer than you'd expect, though still not production-ready. Skyler: The basic photonics integration story goes like this. Pluggable modules — QSFP, OSFP, whatever form factor you're using — those live on the faceplate and have long electrical traces back to the ASIC. CPO, as we just talked about, moves the optical engine onto the package. But CPO still typically uses external light sources — lasers that are co-packaged but not on the silicon photonics die itself, because silicon is really bad at generating light. Silicon is an indirect bandgap semiconductor, and that's a fundamental materials physics problem. Skyler: So what's the solution long-term? You either bond a III-V material like indium phosphide directly to the silicon photonics wafer — which is what most CPO designs do today — or you find a way to grow laser-quality material directly on silicon. Nanolasers are exploring that second path, using quantum dot or other nanostructured gain materials that can potentially be grown on silicon substrates. The advantage would be monolithic integration — one wafer process, lasers and waveguides and detectors all together, no bonding or assembly step. Skyler: The EE Times piece also covers intermediate steps that are happening right now, like photonic wire bonding — using two-photon polymerization to literally 3D print a waveguide in free space between two chips. That sounds wild but it's actually being used in some research and early production contexts. And then there's the chiplet approach, where instead of one monolithic photonic integrated circuit, you break the optical functions into chiplets that can be assembled. That maps well to how the semiconductor industry is thinking about disaggregation generally. Skyler: Why does any of this matter for datacenter architects today? Because the trajectory determines where you invest. If nanolasers on silicon become manufacturable in the next five to eight years — and some researchers think that timeline is realistic — then the cost curve for optical interconnect changes dramatically. Light becomes something that's essentially free at the chip level, and the bottleneck shifts to other parts of the system. For anyone building A I fabric infrastructure today, understanding this roadmap matters for how you think about your optical connectivity vendor relationships over a five-year horizon. Skyler: Story four — Photonics dot com published their March 2026 feature on advanced photonics in next-generation A I datacenters, and it's a good high-level synthesis of where the industry is right now. A lot of what they cover overlaps with what we've already talked about, but they add some useful specifics worth going through. Skyler: One thing the piece emphasizes is the sheer scale of the bandwidth demands that A I training is driving. When you're running a training cluster with tens of thousands of GPUs or accelerators, the all-to-all communication patterns during gradient synchronization are brutal. You need massive bisection bandwidth, you need low latency, and you need it to be power-efficient because the cooling bill is already enormous. Copper can get you there within a rack, but rack-to-rack and pod-to-pod, you're fundamentally in optics territory at the bandwidth densities A I requires. Skyler: The piece gets into some specifics on wavelength division multiplexing inside the datacenter, which is something that used to be purely a long-haul and metro technology. Now you're seeing coarse WDM and dense WDM showing up inside the fabric itself, because it's a way to get more bandwidth through the same fiber infrastructure without having to rewire everything. If you've got a 16-lambda CWDM approach running over existing single-mode fiber, you're multiplying your effective bandwidth by 16 without pulling new fiber. In a datacenter where you're doing a major A I infrastructure build, that's a significant cost lever. Skyler: They also talk about silicon photonics as the manufacturing platform of choice for scaling this up. And I think that's right — the reason silicon photonics has won the platform argument, despite silicon's laser problem we talked about earlier, is that it lets you use CMOS fabs. You can run silicon photonics wafers through the same kind of high-volume, high-yield processes that make your logic chips. That's a massive advantage over compound semiconductor platforms when you're trying to get to tens of millions of ports. Skyler: The piece rounds out with a look at power efficiency, and this is something I want to emphasize because I think it gets underweighted. The target the industry is working toward is something like one picojoule per bit for on-package optical interconnect. Today's pluggable modules are more like ten to twenty picojoules per bit. That's a ten to twenty times improvement. At the scale of a large A I training cluster, that difference in watts is the difference between a datacenter that's viable and one that's not. Photonics isn't just a bandwidth play — it's a power play. Skyler: Okay, last story, and this one's got some real market data behind it. Dell'Oro Group just published their 2025 A I scale-out network fabric numbers, and the headline is striking: Ethernet has more than doubled the size of InfiniBand as the leading fabric for A I scale-out. Not grown faster — more than doubled in absolute revenue terms. That is a significant data point and it's worth unpacking carefully because the narrative implications run in a few different directions. Skyler: First, let's be precise about what we're measuring. Dell'Oro is looking at scale-out fabric specifically — that's the east-west network connecting accelerators within a cluster. InfiniBand has been the traditional champion here because of its RDMA capabilities, low latency, and the tight integration with the GPU ecosystem through Nvidia's stack. Ethernet-based RDMA over converged Ethernet — RoCE — has been the challenger, and for years InfiniBand held a meaningful share advantage in serious A I workloads. Skyler: So what changed? A few things happened simultaneously. Hyperscalers — your Metas, your Googles, your Microsofts — have been aggressively moving to Ethernet-based fabrics, partly for vendor independence, partly because they run their own ASIC programs and can tune the whole stack, and partly because Ethernet just keeps getting faster. 400 gig is table stakes now, 800 gig is rolling out, and the roadmap to 1.6 terabit is real. At those speeds and with RoCE v2 maturation, the performance gap between InfiniBand and Ethernet for many workloads has narrowed significantly. Skyler: There's also the Ultra Ethernet Consortium — U E C — which is specifically trying to formalize Ethernet enhancements for A I workloads. Things like better congestion control, improved transport for all-reduce communication patterns, and host-based offload mechanisms. U E C specs are still evolving — I want to be clear about that, this is not a finished standard — but the fact that AMD, Broadcom, Cisco, HPE, Intel, Meta, and others are all lined up behind it sends a signal to the market that Ethernet is the safe long-term bet for scale-out fabric. Skyler: Now, InfiniBand is not dead. Far from it. Nvidia's NVLink and InfiniBand stack remains deeply integrated and for pure Nvidia-on-Nvidia training clusters, it's still a compelling package. And there are workloads where InfiniBand's latency characteristics are genuinely differentiated. But the market share data from Dell'Oro suggests that at the infrastructure level, Ethernet is becoming the dominant choice — and when hyperscalers write the checks, that's what the market numbers reflect. Skyler: What's interesting strategically is how this interacts with the CPO and photonics stories we talked about earlier. The big Ethernet switch ASICs — your Tomahawk series, your Tofino successors, Cisco Silicon One — these are the platforms that CPO is being designed around for next-generation deployments. So the bet on Ethernet as the scale-out fabric is also, implicitly, a bet on the silicon photonics and CPO ecosystem that's coalescing around those ASICs. These things are more connected than they look on the surface. Skyler: And that's a wrap on today's episode. We covered a lot of ground — from why copper and optics are playing complementary roles rather than fighting each other, to Molex making a smart move into fiber coupling with the Teramount acquisition, to how photonics is pushing all the way toward nanolasers on silicon, to the real power efficiency stakes of optical integration, and finally to Dell'Oro putting hard numbers behind what a lot of us have been sensing: Ethernet is running away with A I scale-out fabric. Line Rate, brought to you by Ada and Zen Semiconductor — Ada helps organizations integrate A I into their workflows in a real and sustainable way, and Zen Semiconductor brings a portfolio of businesses across the silicon industry. If you got value from today's episode, please subscribe and share it with a colleague — word of mouth is genuinely how this show grows. Show notes with links to all five stories are in the episode description. I'll see you next time.