Nokia Tests Ultra Ethernet for AI Data Centre Networks - Data Centre Magazine

Skyler: Hey, welcome to Line Rate — the show tracking the standards, silicon, and companies shaping A I datacenter fabric and high-speed interconnect. I'm Skyler Monroe, and if you live at the bleeding edge of how data moves inside and between the machines training and serving A I, 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 at scale. Today we've got Nokia pushing Ultra Ethernet into production territory, Arista going modular in a big way, Celestial A I pulling in 250 million dollars on the back of photonic fabric, HPE telling us nobody's flinching at rising datacenter equipment costs, and Marvell staking out its own photonic fabric position. A lot to get through — let's go. Skyler: Alright, first up — Nokia. Nokia has been running trials of Ultra Ethernet technology for A I datacenter networks, and this is worth paying attention to because it's one of the cleaner real-world validations we've seen from a major vendor on U E C so far. So for context: the Ultra Ethernet Consortium — U E C — was stood up in 2022 with the explicit goal of building an Ethernet-native transport optimized for A I and HPC workloads. The founding members read like a who's-who — AMD, Arista, Broadcom, Cisco, Eviden, HPE, Intel, Meta, Microsoft — serious weight behind it. The pitch is that you take standard Ethernet as the foundation, which gives you the ecosystem, the tooling, the cost structure, and then you layer in reliability and performance features specifically designed for collective communication patterns — the all-reduce, all-to-all traffic that dominates distributed training. Skyler: Now what Nokia is actually doing here is testing U E C-aligned technology in a datacenter fabric context — validating that this stack actually performs under A I workload conditions. And the thing that makes this interesting is Nokia's position. They're not a hyperscaler, they're not a GPU vendor — they're infrastructure. When an infrastructure player starts testing and publishing on this, it signals that the ecosystem is maturing past the whitepaper phase. U E C specs have been evolving — transport layer, the reliability mechanisms, congestion control — and there's still work to be done on the full specification stack. But vendor-level testing like this is how you go from a consortium document to something that actually ships in a rack. Skyler: The competitive framing here matters too. InfiniBand has long been the default for tightly-coupled A I training workloads — Nvidia's ecosystem essentially owns that space through Mellanox. The bet that U E C is making, and that Nokia is implicitly endorsing by running these trials, is that Ethernet can close the performance gap while delivering the cost and interoperability advantages that InfiniBand structurally cannot. You don't get vendor lock-in with Ethernet the way you do with InfiniBand. For hyperscalers and large enterprises building at scale, that matters enormously. Nokia testing this is a signal that the broader networking industry is getting ready to back that bet. Skyler: And the timing is interesting — you've also got U A Link coming from the AMD-led consortium, which is targeting the scale-up side, the chip-to-chip and node-to-node links within a single training job, while U E C sits more on the scale-out side, the fabric between pods and racks. These aren't necessarily competing — they can be complementary — but the industry is actively sorting out which protocols own which layers, and every piece of real-world testing data points that direction. Nokia's trial results will matter. Watch for more specifics on latency, bandwidth utilization, and congestion behavior as they publish more. Skyler: Next — Arista. The Next Platform had a good deep-dive on Arista's modular switch direction, and this one's got some real substance to it. Arista has been expanding their modular switch portfolio in a way that's trying to thread a needle — you want the density and throughput that scale-out A I fabrics demand, but you also need the flexibility that traditional enterprise and cloud-scale customers expect. Modular chassis designs are how you get there, and Arista is leaning into that hard. Skyler: So what's the actual play here? Arista's modular architecture lets customers slot in different line cards with different speeds and port configurations — you can mix 400G, 800G, and future higher-speed interfaces in a single chassis, which matters a lot when you're building a fabric that's going to evolve over a multi-year lifecycle. You're not ripping and replacing a chassis every time a new SERDES generation lands. The chassis stays, the line cards turn over. That's a meaningful TCO argument, especially when you're talking about the capital intensity of building A I infrastructure. Skyler: The scale-out angle is where it gets interesting for the A I fabric audience specifically. A I training at the cluster level — think hundreds or thousands of GPUs — requires a spine-leaf or multi-stage fabric that can handle all-to-all communication patterns with very consistent low latency. Arista's modular systems are being positioned to anchor the spine layer of these fabrics, with the line card flexibility allowing operators to tune port density and speed to exactly the traffic profile they're running. The EOS software stack — Arista's operating system — is also a differentiator here because it gives you consistent programmability and telemetry across the whole fabric, which is non-trivial when you're trying to debug collective communication performance issues at scale. Skyler: The competitive context: Cisco has its own modular spine play, Juniper is in the mix, and of course you've got merchant silicon vendors like Broadcom with Tomahawk and Jericho families underpinning most of these systems. What Arista is betting on is that systems-level integration, software, and operability will be the differentiator, even when the underlying silicon is similar across vendors. That's been a winning bet for them in cloud networking — the question is whether A I fabric is different enough to shake it up. Given that Arista keeps landing design wins at the hyperscalers, the answer so far seems to be no — software and ops still matter. Skyler: Also worth noting: the scale-out framing in the headline is pointing at the fact that modular doesn't just mean big monolithic chassis anymore. Arista is thinking about how these systems interconnect to scale the fabric horizontally — it's not just one big switch, it's how you stitch many of these together into a coherent fabric. That architectural thinking is what the A I workload demands, and it's good to see it reflected in how they're positioning the product line. Skyler: Okay, this one's a big deal — Celestial A I just closed 250 million dollars in funding to scale its Photonic Fabric technology. Let's unpack what they're actually building, because the term "photonic fabric" is getting thrown around a lot right now and it means different things depending on who's saying it. Skyler: Celestial A I's core thesis is disaggregation — specifically, that the memory bandwidth bottleneck in A I inference and training is best solved by moving data optically between compute and memory rather than over traditional electrical interconnects. Their Photonic Fabric is essentially an optical switching and interconnect layer that lets you pool memory resources across multiple compute elements — CPUs, GPUs, accelerators — without the latency and bandwidth penalties you'd normally pay for going off-chip and off-package. The pitch is that you can build a system where the memory hierarchy is effectively flattened, where an accelerator can reach into a large pool of high-bandwidth memory that's distributed across the system as if it were local. Skyler: Why does this matter? Because the A I scaling laws that have driven the last few years of model growth are running into a hard wall: you can't fit a trillion-parameter model's working set into the on-package memory of a GPU, and moving data over PCIe or even NVLink has latency and bandwidth costs that start to dominate your training and inference efficiency at scale. If Celestial A I can make optical memory interconnect work at the bandwidth and latency that on-package memory provides, that's a genuinely disruptive capability. Big if — but the fact that they're raising 250 million dollars suggests investors think they're past the proof-of-concept phase. Skyler: The funding round itself is a signal. 250 million dollars is serious capital — this isn't seed money for a demo, this is build-out-the-fab-relationships and scale-your-engineering-team money. The photonic interconnect space has been attracting serious capital because the physics genuinely favor light over electrons for certain distance and bandwidth regimes. You avoid skin effect, you avoid the impedance matching challenges, you can wavelength-division multiplex to get enormous aggregate bandwidth in a single fiber or waveguide. Skyler: The challenge — and it's real — is integration. Coupling optical components with the silicon processes that compute happens on is hard. The coupling losses, the packaging complexity, the thermal management — these are not solved problems at production scale. Celestial A I is betting they can get there, and 250 million dollars is a serious runway to prove it out. Watch for product announcements and customer engagements as they deploy this capital. Skyler: Quick one on HPE — The Next Platform reported that HPE is seeing no customer pushback on rising datacenter equipment costs, and honestly, this is a macro signal worth sitting with for a minute. HPE's business spans servers, storage, networking, and high-performance computing, and if their customers are absorbing price increases without pushing back, it tells you something about the demand environment. Skyler: The read I have on this: A I infrastructure buildout is being treated as a strategic imperative by the organizations doing it. When something is strategic — when the board has decided it's how you stay competitive — procurement conversations look very different than when you're buying commodity servers for a routine refresh. You're not squeezing the vendor on margin when you're trying to get your A I cluster online before your competitor does. That's the dynamic HPE is benefiting from right now. Skyler: What's driving equipment cost increases? A few things. High-speed interconnects — 400G and 800G optics — are still not cheap. The ASICs needed for low-latency A I fabric switching carry a price premium. Memory — especially HBM — is tight and expensive. And power density requirements are going up, which means more sophisticated cooling and power delivery infrastructure. All of this adds up to datacenter equipment costs that are genuinely higher than they were three years ago, not just inflation. Skyler: For the industry, the HPE data point is bullish short-term but raises a question: how long does strategic-imperative pricing power last? At some point, A I infrastructure becomes more commoditized, procurement gets more disciplined, and margins compress. That's not this year — and probably not next year for the leading-edge components — but it's the direction of travel. For now, if you're selling into this market, the demand signal is clear. Skyler: The other implication: vendors have pricing power right now, which means they can fund R&D. That matters for interconnect specifically, because the next generation of fabric technology — 1.6 terabit Ethernet, co-packaged optics, photonic switching — requires enormous investment. The margin environment today is what funds the next generation of capability. It's a virtuous cycle while it lasts. Skyler: And lastly — Marvell. So Marvell has been positioning around Photonic Fabric technology, and the timing here is interesting given what we just covered with Celestial A I. Marvell is coming at this from a very different angle — they're a silicon company, they make the ASICs and optical DSPs that go into the interconnect stack, and they're putting the Photonic Fabric framing around their portfolio of optical and photonic silicon IP. Skyler: What Marvell is talking about is their capability to integrate optical functions — specifically, the DSPs, the SerDes, the photonic integration — at the silicon level. They've been building out their optical interconnect silicon portfolio through a combination of internal R&D and acquisitions, and the Photonic Fabric brand is their way of packaging that into a coherent story for A I infrastructure customers. Think of it as Marvell saying: we are the silicon layer that makes photonic interconnect work at scale. Skyler: The specific capability here involves co-packaged optics — where the optical transceivers are integrated much more closely with the switch or compute ASIC, rather than sitting in pluggable form factor modules in a faceplate. Co-packaged optics dramatically reduce the electrical path length between the switch silicon and the optical engine, which reduces power consumption and improves signal integrity. For A I fabric applications — where you're running hundreds of ports at 800G and beyond — the power and density advantages of co-packaged optics are significant. Skyler: Marvell's angle is that they can provide the silicon platform that enables co-packaged optics and photonic integration to scale. They're not building the system — they're providing the components that system builders, hyperscalers, and OEMs integrate. That's a different business model than Celestial A I, which is building a full-stack photonic fabric solution. Both can win — the silicon supplier and the systems integrator can coexist, and in many cases the hyperscalers will want to source the silicon directly and build their own systems. Skyler: What's interesting about both Celestial A I and Marvell making photonic fabric moves in the same news cycle is that it validates the thesis that the optical-electrical interface is where a lot of the next round of A I infrastructure innovation is going to happen. The electrical interconnect inside a server node has largely been optimized — you're not going to get another order-of-magnitude improvement on a copper trace. But the optical layer, from chip to chip, from node to node, from rack to rack — that's still early innings, and the physics allow for substantial improvement. Both companies are betting accordingly. Skyler: That's a wrap on today's Line Rate. Nokia moving U E C into testing, Arista going modular for fabric scale, Celestial A I pulling in 250 million dollars on photonic fabric, HPE telling us nobody's blinking at datacenter costs, and Marvell staking out the photonic silicon layer — a lot of signal in one episode. Line Rate, brought to you by Ada and Zen Semiconductor — Ada helps organizations integrate A I into their workflows in ways that actually stick, and Zen is an umbrella company with a portfolio across the silicon industry. If you got value from this episode, subscribe wherever you get your podcasts, share it with a colleague who needs to be tracking this space, and check the show notes for links to everything we covered today. I'm Skyler — see you next time.