Nokia completes Ultra Ethernet validation of data centre switch products - Telecompaper

Skyler: Welcome to Line Rate — the show tracking the chips, standards, and companies defining how A I infrastructure actually moves data. I'm Skyler Monroe, and if it touches datacenter fabric, high-speed interconnect, or the protocols that stitch racks and datacenters together, we're talking about it here. Huge thanks to our sponsors Ada, who helps organizations actually integrate A I into their workflows in ways that stick, and Zen Semiconductor, an umbrella company with a portfolio of businesses across the silicon industry. Alright — today we've got Nokia putting U E C validation on the board, NVIDIA making the case for network convergence in A I fabric, Celestial A I landing a hundred million dollars to bet on photonic interconnect, and a market sizing report that will make your jaw drop. Let's get into it. Skyler: First up — Nokia. The company just announced it has completed Ultra Ethernet Consortium validation for its datacenter switch products. And look, when I say "Ultra Ethernet," I want to be precise about what that actually means right now, because U E C is still very much a work in progress as a formal spec. The Ultra Ethernet Consortium was stood up in 2023 with a pretty all-star roster — AMD, Arista, Broadcom, Cisco, Eviden, HPE, Intel, Meta, and Microsoft among others — specifically to address the shortcomings of standard Ethernet when you're running A I training workloads at scale. The core problem they're solving is that classic Ethernet was designed for bursty, latency-tolerant traffic. A I collective communication — your all-reduce operations, your gradient synchronization — that's a different beast. You need deterministic low latency, you need to handle congestion without collapsing throughput, and you need transport semantics that actually map onto how these A I frameworks think about communication. So U E C is defining a new transport layer — sitting above the physical and link layers, building on top of Ethernet's ubiquity but adding things like out-of-order packet handling, better multipathing, and hardware-friendly transport that doesn't need a full TCP stack. Now, Nokia completing "validation" here — this is important to unpack. U E C's interoperability testing is still maturing. What this represents is Nokia running its datacenter switch portfolio through the consortium's current conformance and interoperability frameworks and coming out the other side with a stamp. That matters commercially because hyperscalers and cloud builders evaluating their fabric options want to see exactly this — a major networking vendor committing to the spec and proving their silicon behaves correctly against it. For Nokia specifically, this is a meaningful move. They're not historically the first name you think of for datacenter switching — that's Arista, Cisco, Broadcom-based white box — but they've been pushing hard into the datacenter with their IP fabric and optical transport heritage. Getting U E C validation early is a signal to the market: we're serious about A I networking, and we're not going to be caught flat-footed the way some vendors were when InfiniBand started eating their lunch in the A I training segment. The broader competitive implication is real. U E C, if the spec matures the way the consortium intends, could be a genuine challenger to InfiniBand for A I fabric — and every switch vendor that gets validation early builds a reference point for the procurement conversations that are going to happen over the next eighteen months. Watch this space closely. Skyler: Alright, story two. NVIDIA dropped a technical blog post this week on A I fabric resiliency and why network convergence matters — and honestly, this is worth pulling apart because it's not just marketing copy. There's real substance here about how NVIDIA thinks about the fabric layer in their A I infrastructure stack. The core argument NVIDIA is making is that as you scale A I training and inference clusters, the network is no longer a commodity layer you can treat as an afterthought. Failures in the fabric — link flaps, switch errors, congestion events — have outsized impact on training job throughput because these workloads are so tightly coupled. A single slow link in an all-reduce can drag the whole collective down. That's not new insight, but the way they're framing it around "convergence" is interesting. What they mean by convergence here is the idea of collapsing the separate networks that have traditionally served different functions in the datacenter — storage, inter-GPU communication, management traffic — onto a single, unified, high-performance fabric. The pitch is that a converged fabric simplifies operations, reduces the number of distinct network domains you have to manage and monitor, and when it's designed right, actually improves resiliency because you have more paths and more flexibility to reroute around failures. Now, when NVIDIA talks about this, they're obviously not operating in a vacuum — they're talking about their Spectrum-X platform and their broader NVLink and InfiniBand ecosystem. The technical blog goes into failure detection, fast failover, and how their approach to congestion control — specifically their A I Optimized Ethernet work — handles the kinds of traffic patterns A I workloads generate. The resiliency angle is really pointed at a real pain point. Operators running large A I clusters will tell you that job interruptions from network events are one of the top causes of wasted GPU time. When you're paying tens of thousands of dollars per GPU-hour, a training run that crashes because of a fabric hiccup and has to checkpoint-and-restart is brutal. So NVIDIA is positioning their converged fabric story as a way to reduce that blast radius — better monitoring, faster detection, smarter rerouting. The competitive context here is that this blog lands right as alternatives like U E C and UALink are developing, and Arista and Cisco are making their own pitches to hyperscalers about Ethernet-based A I fabric. NVIDIA has an obvious advantage in the NVLink domain — that's proprietary and tightly integrated with their GPU architecture — but at the scale-out layer, the Ethernet versus InfiniBand debate is very much alive. This blog is part of NVIDIA making the case that their end-to-end fabric story, not just the GPUs, is the right answer. It's a smart move to own the narrative around resiliency when that's exactly what operators are losing sleep over. Skyler: Story three — this one is genuinely exciting if you follow the optical interconnect space. Celestial A I just closed a hundred million dollar Series B to accelerate what they're calling "photonic fabric" for datacenters. Let's talk about what that actually means and why a nine-figure Series B is landing here right now. Celestial A I's approach is built around what they call the Optical Multidie Interconnect Engine — basically, they're using photonics not just for rack-to-rack or datacenter-to-datacenter links, which is where optical has traditionally lived, but pushing it much closer to the compute — think die-to-die and chiplet-to-chiplet interconnect using photonics rather than electrical signaling. The fundamental physics argument is compelling: as you try to move more data over shorter distances at higher speeds, electrical interconnect runs into serious bandwidth density and power consumption walls. Copper traces on a package or between chiplets dissipate a lot of energy as heat, and you hit signal integrity limits that force you into more complex equalization, which burns more power. Photonics sidesteps a lot of those problems — light doesn't care about impedance matching the way electrons do, and you can pack enormous bandwidth into a single fiber or waveguide. The reason this matters for A I fabric specifically is that the compute-to-memory bandwidth bottleneck and the chip-to-chip communication bottleneck are increasingly the limiting factors in A I system performance. You can have the fastest GPU in the world, but if you can't feed it data fast enough or synchronize across chips fast enough, you're leaving performance on the floor. Celestial's pitch is that photonic fabric can move that bottleneck significantly. Now, a hundred million dollar Series B is a meaningful signal that institutional investors see this as more than a science project. The A I infrastructure buildout has created a very unusual environment where the capital appetite for interconnect innovation is genuinely large. The hyperscalers are spending billions on datacenters, and even a small improvement in interconnect efficiency translates to massive TCO savings at their scale — so the customer pull is real. The competitive landscape here is fascinating. You've got established players like II-VI — now Coherent — and Lumentum on the optical components side, Intel with their silicon photonics work, Ayar Labs doing optical I/O chiplets, and a cohort of startups including Celestial, all racing toward what is essentially the same vision: get photonics as close to the compute as possible. The question is who can get to the cost and integration points that make it manufacturable at datacenter scale, not just impressive in a lab. Celestial's hundred million dollars buys them time and engineering depth to push toward that. Worth watching very closely over the next two years. Skyler: Story four — this one is more of a market context story, but I think it's worth spending a few minutes on because the numbers are big and the framing tells you something about where the industry thinks this is all headed. MarketsandMarkets put out their datacenter networking market report covering 2025 through 2031, and the headline numbers are substantial — we're talking about a market that's projected to grow significantly over that window, driven overwhelmingly by A I infrastructure buildout. Now, I'll be honest with you — market research reports like this come with the usual caveats. The methodology matters, the assumptions about adoption curves matter, and frankly some of these firms have a history of numbers that flatter the narrative. So take the specific figures with appropriate skepticism. But the directional story is consistent with what you see from every other data point in this space: datacenter networking is in a structural growth cycle unlike anything since the early cloud era. What's driving it? A few things. First, A I training and inference clusters are network-intensive in ways that traditional cloud workloads simply aren't. We've already talked about this — tight collective communication, sensitivity to latency, enormous east-west bandwidth requirements. That means more ports, higher speeds, and more sophisticated switching. Second, the speed upgrade cycle is accelerating. We're in the middle of a transition from 400 gig to 800 gig at the spine, with 1.6 terabit starting to show up in roadmaps and early deployments. Every speed jump generates a significant capex cycle as operators upgrade their fabric. Third, the geographic diversity of A I infrastructure buildout is real — it's not just the US hyperscalers anymore. You're seeing serious investment in Europe, in Asia, in the Middle East. That expands the total addressable market. The report segments by offering type, technology, and geography, and unsurprisingly the Ethernet and InfiniBand segments and the switching silicon segment are where the action is. The implication for the companies we cover on this show — switch vendors, silicon vendors, optics companies, standards bodies — is that the rising tide is real, but so is the competition. When the market is this large and growing this fast, it attracts capital and new entrants. The incumbents who get complacent get disrupted. The startups who can actually deliver at scale get acquired or go public. It's a target-rich environment, and the next five years are going to be genuinely chaotic in the best possible way. Skyler: And then there's story five, which is a bit of a left turn from our usual beat, but bear with me — Blocks and Files flagged something from a company called UnicornIQ, and specifically their product called Vera, which is a data fact checker designed for sales use cases and sales chatbots. Now, why is this on a datacenter interconnect show? Fair question. The direct connection to fabric is thin, but I think it's worth a minute because it represents something broader that's happening across the A I tooling landscape that our audience should be aware of. As A I gets deployed into enterprise workflows — and it's getting deployed fast — the accuracy problem becomes critical. Sales teams using A I chatbots or automated tools that hallucinate product specs, pricing details, or competitive claims is not a theoretical risk. It's happening. Vera is positioned as a guardrail layer that checks the outputs of A I systems against verified data sources before those outputs reach customers or sales reps. That's actually a non-trivial technical problem — you need fast retrieval, reliable grounding, and low enough latency that you're not killing the user experience. The reason I think this is relevant to our audience specifically is that as our industry — silicon vendors, networking companies, standards bodies — deploys A I into their own customer-facing tools, the accuracy requirements are extremely high. A chatbot that gets a spec wrong for a 400 gig transceiver or misstates the latency characteristics of a switch is not just embarrassing, it's a sales liability. So the category of tools that Vera represents is going to matter to the go-to-market side of the companies building the infrastructure we cover. It's early, the company is small, but the problem they're solving is real and the timing is right. Keep an eye on how this category develops — fact-checking and grounding infrastructure for enterprise A I is going to be its own meaningful market. Skyler: Alright, that's a wrap on today's Line Rate. We covered Nokia's U E C validation — a meaningful early-mover signal in the Ultra Ethernet race. We dug into NVIDIA's case for converged, resilient A I fabric and why network uptime is increasingly a competitive differentiator. Celestial A I's hundred million dollar Series B showed that photonic interconnect is attracting serious capital. We contextualized the datacenter networking market report — big numbers, real structural growth, intense competition. And we took a quick look at UnicornIQ's Vera as a signal of where enterprise A I tooling is heading. Thanks as always to our sponsors — Line Rate, brought to you by Ada and Zen Semiconductor. Ada helps organizations integrate A I into their workflows in ways that actually deliver value. If you found this useful, subscribe wherever you get your podcasts, share it with a colleague who's deep in the stack, and check the show notes for links to everything we covered today. I'm Skyler Monroe — see you next time.