Why Ethernet Matters for AI Networking - Nokia

Skyler: Hey, welcome to Line Rate — the show about datacenter fabric and A I high-speed interconnect. I'm Skyler Monroe, and if you care about the chips, cables, and standards moving data between GPUs, across racks, and between datacenters, you're in the right place. Quick thanks to our sponsors: Ada, who helps businesses actually integrate A I into their real workflows, and Zen Semiconductor, an umbrella company with a portfolio across the silicon industry — including Ada for A I deployment at scale. Alright, today we've got a lot of photons to chase. We're talking about why Ethernet keeps winning the A I fabric argument, a genuinely staggering market projection for optical interconnects, a brand new industry consortium forming around optical connectivity, and then we're going to dig into what co-packaged optics is actually doing to the datacenter supply chain. Let's get into it. Skyler: So the first story comes out of Nokia, and it's a piece on why Ethernet matters for A I networking. Now, I know some of you just rolled your eyes a little — like, do we really need another "Ethernet is great" explainer in 2025? But stick with me, because the framing here is actually useful, especially as Ultra Ethernet Consortium and U A Link are both out there competing for mindshare in the A I fabric space, and the lines between them are still pretty blurry for a lot of people. Skyler: So the core argument Nokia is making is that Ethernet isn't just a legacy protocol being awkwardly retrofitted for A I — it's genuinely the right architectural foundation for large-scale A I training and inference fabrics. And the reason that argument needs to be made explicitly is because InfiniBand has spent years being the default answer when someone says "I need low-latency, high-bandwidth GPU interconnect." Nvidia built an ecosystem around it, and it worked. It works really well inside a pod, inside a rack, inside a cluster you fully control. But as clusters get bigger, as you start spanning racks, rows, multiple datacenters, InfiniBand's operational complexity becomes a real burden. Ethernet, for all its imperfections, is something operators already know how to run at scale. Skyler: What Nokia is pointing to — and this aligns with what Ultra Ethernet Consortium has been working on — is that the missing piece for Ethernet in A I workloads has always been the transport behavior. Standard Ethernet doesn't give you the congestion control, the ordered delivery, the latency predictability you need when you've got a hundred thousand GPUs doing synchronized all-reduce operations. One slow link and your whole training job stalls. That's the ECMP problem, that's the packet reordering problem, that's what UEC is trying to solve with new transport specs layered on top of standard Ethernet PHY. Skyler: And this is where the timing matters. U E C — the Ultra Ethernet Consortium — has been releasing spec work, and while a lot of it is still evolving, the direction is clear: they want to give you InfiniBand-class transport behavior on Ethernet infrastructure. U A Link, which is the consortium backed by AMD, Intel, Google, Meta, and others, is taking a slightly different approach — it's focused on the scale-up side, the intra-node and near-node fabric, not necessarily the full datacenter-spanning scale-out. So these two aren't really competing with each other directly. U A Link is more about replacing NVLink in a multi-vendor world. U E C is about the broader datacenter fabric. Skyler: Nokia's piece is essentially making the case that this trajectory — Ethernet inheriting the A I fabric role — is inevitable, and that the ecosystem investment in silicon, in tooling, in operations, is just too large for anything else to displace it at scale. I think that's largely right. The interesting tension is whether UEC specs land in time to matter for the current GPU buildout cycle, or whether hyperscalers end up rolling their own transport solutions and UEC becomes a catch-up standardization effort. That's a real risk. But the momentum is there, and Nokia clearly wants to be positioned as the infrastructure vendor who saw it coming. Skyler: Next up — and this one is a number that genuinely made me do a double take — projections out from new research suggest the A I datacenter optical interconnect market is going to hit 144 billion dollars by 2030. That is a more than ten-fold increase from 2024 figures. Ten times. In six years. That is not a modest growth projection. That is an industry transformation. Skyler: The story was covered by Tom's Hardware, but the underlying projection is from market research, and the headline number is backed up by some specific structural claims worth unpacking. The biggest one is that silicon photonics is expected to account for nearly two-thirds of that revenue — and the driver is co-packaged optics, or CPO. So let's talk about why those two things are connected. Skyler: Right now, the dominant optical module form factor in datacenters is pluggable — you've got your QSFP-DD, your OSFP, your 800G and 1.6T modules sitting in a faceplate cage, connected to your ASIC via a PCB trace. That works, but every millimeter of that electrical trace between the switch ASIC and the optical module burns power and adds latency. At 800 gig per port that's manageable. At 3.2 terabit per port, which is where high-radix spine switches are heading, the SerDes power and signal integrity on that electrical segment becomes the limiting factor. You just can't push that much bandwidth over copper traces efficiently at the distances involved. Skyler: Co-packaged optics solves this by putting the optical engine right next to — or in some implementations, directly integrated with — the switch ASIC in the same package. You're eliminating that PCB segment entirely. The optical transceiver dies are in the same package as the switching silicon. The electrical path goes from centimeters to millimeters or less. That's a massive power efficiency gain. Some estimates put the power savings at 30 to 50 percent compared to pluggable at equivalent bandwidth density. Skyler: Silicon photonics is the enabling technology here because you can't package traditional III-V optical components — your indium phosphide lasers, your lithium niobate modulators — directly with a CMOS switch ASIC at scale. The manufacturing processes aren't compatible, the thermal profiles don't match, and the packaging complexity is brutal. Silicon photonics lets you build the modulator, the waveguide, the photodetector on a silicon substrate using processes that are at least compatible with, if not identical to, standard CMOS flows. You can then flip-chip bond a laser die and get your light source. It's still not trivial, but it's the approach that scales. Skyler: So when you see that two-thirds figure for silicon photonics, that's really a proxy for CPO adoption. If CPO hits the inflection point people are expecting — and the new consortium announcement we're going to talk about next suggests the ecosystem is trying hard to make that happen — then silicon photonics captures the majority of the optical revenue because it's the only manufacturable path to CPO at datacenter scale. The 144 billion dollar number is big, but the structural logic behind it is pretty sound. Skyler: Okay, story three. A new organization just formed to support optical interconnects for A I datacenter applications. The announcement came out on Business Wire, and this is the kind of consortium news that can sound like a press release and nothing else — but I think this one is worth paying attention to for a specific reason. Skyler: The group is formed by "leading optical connectivity solutions providers" — the announcement is pretty high-level on who exactly is in it, which is typical for a formation announcement. But the stated mission is to accelerate standardization and interoperability for optical interconnects specifically in A I datacenter applications. And the timing is not an accident. This is forming right as the industry is trying to navigate the transition from pluggable optics to CPO, and there are genuinely unresolved questions about how CPO modules get specified, tested, and qualified across different switch ASIC vendors and optical engine suppliers. Skyler: Here's the problem they're trying to solve. Right now, if you're a hyperscaler and you want to deploy CPO-based switches, you're looking at a relatively closed ecosystem. Broadcom is working on CPO integration with their Tomahawk and Jericho lines. Marvell has CPO work ongoing. Intel has their own silicon photonics roadmap and CPO aspirations. But the interface between the optical engine and the switch ASIC — what's the electrical spec, what's the thermal management contract, what are the testing and qualification criteria — none of that is standardized. Every vendor is rolling their own. Skyler: That creates serious supply chain risk for buyers. If you qualify a CPO solution from Vendor A and they have a supply disruption, you can't just swap in Vendor B's optical engine. The interfaces aren't interoperable. You're back to a single-vendor lock-in situation, which hyperscalers hate. The whole point of an open standard is to create a competitive supplier ecosystem, which drives down prices and de-risks the supply chain. Skyler: So a consortium forming around this problem is genuinely useful work, assuming they can get the right companies at the table and actually produce specifications that vendors commit to. The risk — and this is the risk with every consortium announcement — is that it becomes a lowest-common-denominator spec that satisfies everyone because it's vague enough to not actually constrain anyone. That's not a standard, that's a marketing document. The test will be whether they produce interface specs with enough precision that you can qualify a switch ASIC from one vendor against an optical engine from another vendor and have confidence it works. Skyler: We've also got OIF, we've got CXL Consortium, we've got IEEE, we've got Ultra Ethernet Consortium all doing adjacent work. The optical interconnect space does not have a shortage of consortia. What it has a shortage of is binding, implemented specifications. So color me cautiously optimistic on this one. The problem is real, the motivation is there, and the market pressure to solve it is enormous. But I'll be watching to see if this produces actual technical work or just white papers. Skyler: Story four comes from Photonics Spectra, a March 2026 feature titled "Integrated Optics: Breaking the Bandwidth Bottleneck for Hyperscale A I." This is a deeper technical piece, and it's framing integrated optics — which is essentially the umbrella term covering silicon photonics, CPO, and on-chip optical interconnects — as the answer to a scaling wall that hyperscalers are running into right now. Skyler: The bandwidth bottleneck framing is important. When people talk about A I training cluster performance, they often focus on compute — teraflops of the GPU, memory bandwidth of the HBM. But for large distributed training jobs, the bottleneck shifts to the fabric. Your all-reduce operation across a cluster is only as fast as your slowest network link, or more precisely, your aggregate bisection bandwidth. And as cluster sizes have gone from thousands to hundreds of thousands of GPUs, the interconnect demand has scaled non-linearly. You need more bandwidth, yes, but you also need lower latency and better congestion characteristics, because the synchronization overhead in large all-reduce operations is exquisitely sensitive to tail latency. Skyler: The Photonics Spectra piece walks through why integrated optics specifically addresses this in a way that pluggable optics or copper DAC cables can't. It's not just the bandwidth per port — it's the bandwidth density. How many bits per second can you move through a given area of switch faceplate, or through a given volume of rack space? Pluggable modules are hitting fundamental physical limits. You can only get so many OSFP cages on a faceplate before you run out of room, and each cage has thermal constraints on how much power the module can dissipate. Skyler: Integrated optics, and particularly CPO, busts through that constraint by moving the optical engine off the faceplate and into the package. You get dramatically higher port density per unit of switch silicon because you're not limited by the faceplate cage count. And you get better thermal management because the optical engine can be designed for the thermal environment of the package rather than the thermal environment of a faceplate cage in a live switch. Skyler: The feature also touches on where the technology is actually at versus where it needs to be. Silicon photonics insertion loss is still higher than you'd like — every coupler, every bend in a waveguide costs you optical power, and that power comes from your laser. Laser integration is still the hard part. You can build a fantastic silicon photonics modulator and waveguide structure on CMOS, but silicon is a lousy light emitter. You need a III-V laser source — indium phosphide or gallium arsenide — and getting that coupled efficiently into your silicon waveguide at manufacturing scale, with high yield, is an unsolved mass production problem. Not unsolved in the lab — there are teams doing it. Unsolved at the volumes that hyperscale deployment requires. Skyler: The article is bullish on the trajectory, and I think the bullishness is justified, but the timeline risk is real. The market projection we talked about — 144 billion dollars by 2030 — assumes CPO adoption accelerates pretty significantly in the next two to three years. If laser integration yield issues delay CPO deployments by two years, those numbers look very different. Worth keeping in mind. Skyler: Last story today, and this one is really interesting because it's asking a question that most of the industry is dancing around but not directly answering: is copper out and fiber in? And what does that mean for supply chain profit distribution in the CPO era? Skyler: The piece is from Futu — a Chinese financial media platform — and the framing is explicitly about where the money flows in a world where CPO is the dominant interconnect form factor inside the datacenter. This is the supply chain economics angle, and it's one that doesn't get talked about enough in the technical press. Skyler: So let's start with the copper question. Inside a rack, copper DAC cables — direct attach copper — still make a ton of economic sense. They're cheap, they're low latency, they work great at short distances. Nobody is ripping out copper DAC for rack-level connectivity. The fight is at the rack-to-rack and row-to-row level, which is the ToR-to-spine link, and at the longer distances in leaf-spine and fat-tree fabrics. As port rates move from 400 gig to 800 gig to 1.6T, the reach of copper DAC drops. At 800 gig, you're looking at maybe a couple of meters for passive copper. Active copper cables extend that a bit, but active electrical cables have their own power and cost issues at those data rates. Optical is the only thing that gives you the reach you need at high bandwidth, and that transition is already happening. Skyler: So "is copper out" depends heavily on where in the fabric you're talking about. At the rack scale, copper stays. At the datacenter fabric scale — ToR uplinks, spine links, inter-rack connections — fiber is winning and will continue to win. The question is just which flavor of fiber. Skyler: Now for the profit distribution question — this is where CPO gets really disruptive. In the pluggable optics world, the supply chain is pretty well defined. You've got the switch ASIC vendor, the switch OEM, and the optics module vendor — your Coherents, your II-VIs, your Innolight — each capturing a slice of the bill of materials. The optics module is a discrete product with its own margin structure. Hyperscalers are paying 200, 300, 400 dollars per port for high-end pluggable modules. Skyler: In the CPO world, the optical engine is integrated into the switch package. The optical engine vendor — which might be the ASIC vendor themselves, or a specialized silicon photonics company like Ayar Labs or Celestial AI or Intel's photonics group — captures value at a different point in the stack. The switch OEM might be less relevant in some architectures because hyperscalers are buying direct from ASIC vendors or doing their own integration. The traditional pluggable optics module vendors either need to pivot to making CPO optical engines, or they get commoditized or displaced. Skyler: This is the existential question facing companies like Coherent and Lumentum and II-VI right now. They've built enormous businesses on pluggable transceiver revenue. CPO doesn't eliminate the need for their core competencies — laser fabrication, optical packaging, photonic integration — but it does restructure where those competencies get applied and who captures the margin. If the A I hyperscalers are buying CPO-integrated switch ASICs from Broadcom or Marvell, and Broadcom is sourcing their optical engines from a silicon photonics foundry, the traditional module vendor is several steps removed from the revenue. Skyler: The smart play for incumbent optics companies is to become the optical engine supplier for CPO programs — which means developing silicon photonics capability and the packaging know-how to co-integrate with switch ASICs. Some of them are doing exactly that. But it requires significant investment and a different engineering culture than building pluggable modules. The supply chain is being restructured in real time, and the companies that read that transition correctly are going to capture a lot of that 144 billion dollar pie. The ones that don't will be fighting for the residual pluggable business in markets that aren't A I hyperscale. Skyler: Alright, that's what I've got for today. Five stories, a lot of photons, and a supply chain