Why Ethernet Matters for AI Networking - Nokia

Skyler: Welcome to Line Rate — the show about datacenter fabric, AI high-speed interconnect, and the standards, silicon, and companies racing to keep up with the data. I'm Skyler Monroe, and if you care about what happens between the chips, between the racks, and between the buildings, this is your show. Big 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, including Ada for A I deployment. Alright, today we've got Ethernet making its case for A I networking, a 125 million dollar optical switching raise with some heavyweight names attached, Ultra Ethernet getting the explainer treatment it deserves, a data center in Norway that's having a very bad time, and a UK council thinking creatively about heat. Let's get into it. Skyler: So Nokia put out a piece this week on why Ethernet matters for A I networking, and look — on the surface this feels like a "water is wet" headline. Ethernet is everywhere. But there's actually a real argument being made here, and it's worth unpacking because the underlying question is genuinely contested. The context is that for the last several years, the dominant interconnect for tightly coupled A I training clusters has been InfiniBand. Nvidia's HPCX stack, NCCL tuned for IB, the whole ecosystem. And for good reason — InfiniBand gives you extremely low latency, hardware-based congestion control, and a mature software stack for collective operations. The problem is it's largely a Nvidia-controlled ecosystem, and as hyperscalers and large enterprises try to build A I infrastructure that isn't completely captive to one vendor, they start looking hard at Ethernet as an alternative. Skyler: Nokia's argument, and this tracks with where the broader industry is heading, is that Ethernet's ubiquity, its cost structure, and critically the work being done in the Ultra Ethernet Consortium gives it a credible path to closing the gap on the performance metrics that matter for A I workloads. We're talking about congestion management, packet ordering, and efficient handling of the all-to-all traffic patterns that dominate large-scale training runs. These are the areas where vanilla Ethernet has historically fallen down relative to InfiniBand. The U E C is directly attacking those gaps. Things like receiver-side packet reordering, adaptive routing across multiple paths, and new transport mechanisms that don't require the overhead of standard TCP for these kinds of workloads. Skyler: Now, Nokia has real skin in the game here — they sell high-end routing and switching that feeds into datacenter fabrics, and more Ethernet in A I clusters is good for that business. So you should read this with that lens. But the technical substance isn't wrong. The question isn't really whether Ethernet can work for A I fabric — it's whether the U E C can deliver a spec tight enough and fast enough that silicon vendors actually converge on it, and whether the software stack catches up to what IB users have today. And we'll talk more about U E C in a few minutes because there was a SlashGear piece this week that's worth getting into. But the Nokia framing is useful — Ethernet for A I isn't a compromise story, it's becoming a legitimate architecture story. Whether that plays out in 2025 or 2027 is the open question. Skyler: Okay, let's talk about Salience Labs, because 125 million dollars Series B with AMD and Intel in the round is not a quiet funding announcement. Salience Labs is building optical circuit switching technology — OCS — and this raise is specifically to support continued development and deployment of that. So let me give you the quick primer if you haven't been tracking this space. Traditional datacenter switching is electrical. Data comes in on a fiber, gets converted to an electrical signal, gets switched through silicon, gets converted back to optical for the next hop. Every one of those conversions costs you power and latency. Optical circuit switching cuts out some or all of that conversion — you're routing the light itself, not the electrical representation of it. Skyler: The appeal for A I fabric is real. A I training clusters move enormous volumes of data between GPUs and accelerators, and the traffic patterns — especially for things like all-reduce and all-to-all collectives — are somewhat predictable. They're bursty in a structured way. That's actually a good fit for circuit switching, where you pre-establish a path and then send a lot of data down it, as opposed to packet switching where every packet makes an independent routing decision. Google has been using OCS in their datacenter interconnect for years, and they've published extensively on it. The idea that you can dynamically reconfigure an optical circuit fabric to match the traffic demand of a training run is compelling — lower power, lower latency on the data path, potentially much higher bandwidth efficiency. Skyler: What makes this raise interesting beyond the dollar figure is who's in it. AMD and Intel are not passive observers here. Both of them are building A I accelerators — MI300 and Gaudi lineages respectively — and both of them are deeply interested in interconnect architectures that don't require you to buy into Nvidia's NVLink ecosystem. If Salience Labs can deliver optical switching that plugs into an open interconnect architecture — and the U E C and U A Link work are relevant context here — then AMD and Intel have a potential path to offering a complete A I cluster solution that competes on both the accelerator and the fabric layer. That's strategic, not just financial. We don't know the terms of the round or the ownership structure, but strategic investors at this stage don't write 125 million dollar checks without a product roadmap they believe in. Skyler: The challenge for OCS, and it's a real one, is reconfiguration time. Electrical packet switches can make forwarding decisions in nanoseconds. Optical circuit switches — even fast ones using MEMS or silicon photonics approaches — operate on timescales of microseconds to milliseconds for reconfiguration. For truly dynamic traffic that changes on short timescales, that's a problem. But if you're running structured A I workloads where the communication pattern is known ahead of time, you can schedule the circuits and hide that reconfiguration latency. Salience is clearly betting that the A I workload domain is structured enough to make that work. With 125 million dollars and two major chip vendors backing them, they're going to get a real shot at proving that out. Skyler: Alright, quick sidebar that I find genuinely interesting from a systems perspective — there's a story out of the UK this week about a council in Oxfordshire launching a heat network feasibility study for proposed data centers near Didcot Power Station. Nordic Energy is doing the study, results in March. And look, I know this isn't fabric or switching, but bear with me for a second because it's actually a nice window into how the physical constraints of data centers are starting to shape where they get built and how they get designed. Skyler: Data centers run hot. Especially A I data centers — you're looking at rack densities that can exceed 100 kilowatts per rack with high-end GPU clusters, and all of that heat has to go somewhere. Traditionally it goes into the air or into cooling towers. But heat networks — where you capture that waste heat and pipe it to nearby buildings for space heating — are genuinely attractive in a world where energy efficiency and sustainability are real operational and regulatory concerns. Didcot is interesting because it's an existing power infrastructure site with grid connection, which is one of the biggest bottlenecks for new data center development in the UK right now. You can't build a data center if you can't get the power, and grid connection queues in the UK are measured in years, not months. Skyler: The feasibility study angle is the honest part of this story — no one's claiming this is built or committed. But the fact that local government is proactively looking at how data center waste heat integrates into community infrastructure is a sign of how seriously the physical footprint of A I compute is being taken. If you're designing a data center campus and you can offset heating costs for surrounding buildings with your waste heat, that's a real economic and political win. It probably won't show up in your fabric architecture decisions, but it might determine whether you can get planning permission at all. And planning permission is its own kind of interconnect bottleneck. Skyler: Okay, Ultra Ethernet — let's actually dig into this because SlashGear ran a piece this week asking "what is Ultra Ethernet and who is it for," and it's a fair question. I want to give you the version that's actually useful for people who work in this space, because the consumer tech framing doesn't quite capture what's technically at stake. Skyler: Ultra Ethernet is the work product of the Ultra Ethernet Consortium, which launched in 2023 with founding members including AMD, Arista, Broadcom, Cisco, Eviden, HPE, Intel, Meta, and Microsoft. Notably absent: Nvidia and Juniper at launch, though the consortium has been growing. The U E C's mandate is to define Ethernet-based transport specifications optimized specifically for A I and HPC workloads. Not general-purpose data center traffic, not cloud storage — the specific, brutal demands of large-scale machine learning training where you have thousands of accelerators doing collective operations across a low-latency fabric. Skyler: So what are they actually specifying? There are a few key areas. First, congestion control — traditional Ethernet's mechanisms like ECN and PFC were not designed for the traffic patterns you see in A I training. You get massive incast events where hundreds of senders all try to hit a single receiver simultaneously, and standard congestion control falls apart. U E C is defining new mechanisms here. Second, transport — there's work on a new transport layer that sits below the socket API and provides the kind of reliable, ordered, low-latency delivery that A I frameworks need without the overhead of full TCP. Third, routing — adaptive routing across multiple equal-cost paths, with the intelligence to steer flows away from congestion in real time. Skyler: The "who is it for" question is real. InfiniBand users who are happy with their Nvidia stack don't need U E C. But anyone building A I infrastructure who wants vendor optionality — hyperscalers who already operate massive Ethernet networks, enterprises who don't want to run two parallel fabric architectures, accelerator vendors who aren't Nvidia — U E C is directly relevant. The spec is still evolving. As of now the U E C has released early specification documents but we are not yet at a point where you have broadly available silicon that fully implements the whole stack. That's the honest status. You'll see silicon vendors claiming U E C compatibility and alignment, but the full spec convergence and interoperability testing is ongoing work. Skyler: One thing worth clarifying because it comes up — U E C and U A Link are different things that sometimes get conflated. U A Link, from the U A Link Consortium, is about the scale-up interconnect — the direct chip-to-chip links within a single server node or a tightly coupled pod, the domain where NVLink operates. U E C is about the scale-out fabric — the switched network between nodes and racks. Both are important, they address different layers of the A I interconnect stack, and you'll probably want both if you're building a serious alternative to the Nvidia ecosystem. But they're separate efforts with different member compositions and different technical scope. If you see a headline that conflates them, that's a sign the author hasn't fully tracked the space. Skyler: Last story, and this one is a cautionary tale. Nscale's Glomfjord data center in Norway has been delayed after the site suffered leaks in its cooling system and a fire. This is particularly unfortunate timing because Nscale is planning an IPO in the near term. Glomfjord is a significant site — it's in a fjord location in Norway that offers natural cooling advantages, which is a big part of the pitch for Nordic data center development generally. Cold ambient temperatures, access to renewable hydropower, the whole package. Skyler: The specific failure modes here — cooling leaks and a fire — are both serious in different ways. Cooling system failures in a high-density A I data center are especially consequential because of those rack densities we talked about. If your cooling fails in a facility running 50 to 100 kilowatt racks, you have minutes, not hours, before hardware starts taking damage. The fire is a separate concern and we don't have details on what caused it or how contained it was, but any fire event at a data center construction or early operation stage requires a full investigation and typically a significant reset of timelines. Skyler: The IPO angle makes this awkward. When you're trying to go public, operational reliability and execution confidence are core parts of the investor narrative. A delay on a flagship site due to infrastructure failures — not geopolitical issues, not supply chain, but a cooling leak and a fire — is the kind of thing that shows up in prospectus risk factors and gives underwriters pause. Nscale's overall business is GPU cloud infrastructure and they've been growing aggressively in the European market. Glomfjord is a real site with real strategic value because of its renewable energy profile and cooling economics. But you have to actually get the facility built and operating before that value materializes, and right now they've taken a meaningful setback. Skyler: The broader lesson here — and it's one that anyone speccing out a data center build should sit with — is that the physical plant is as much a technical challenge as the network or the silicon. You can have the best fabric architecture in the world, you can have great GPU density planned out, but if your cooling system leaks or your facility has a fire, none of that matters. Commissioning a high-density data center is genuinely hard engineering, and the failure modes are costly when they happen at scale. Skyler: Alright, that's the show for this week. Ethernet making its A I case, optical switching raising serious money with serious backers, Ultra Ethernet getting some overdue explainer coverage, a UK council thinking cleverly about waste heat, and a Norwegian data center that's had a rough few weeks. As always the show notes will have links to the underlying sources so you can dig into any of these threads further. Line Rate, brought to you by Ada and Zen Semiconductor — Ada helps organizations integrate A I into their workflows in ways that actually work, and Zen Semiconductor is an umbrella company with a portfolio of businesses across the silicon industry. If you got value from today's episode, subscribe wherever you get your podcasts, leave a rating if you're feeling generous, and share it with a colleague who cares about where the data goes. I'll see you next time on Line Rate.