Leading Optical Connectivity Solutions Providers form New Organization to Support Optical Interconnects for AI Data Center Applications - Business Wire

Skyler: Welcome to Line Rate — the show about datacenter fabric, A I interconnect, and the hardware moving data at the speed that actually matters. I'm Skyler Monroe, and if you care about what's happening inside the rack, between the racks, and across the datacenter, you're in the right place. Big thanks to our sponsors — Ada, who helps companies 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 deploying A I at scale. Today we've got a packed show — a new optics consortium just stood up, Synopsys is shipping U E C and U A Link IP, Cisco is making a play for unified fabric operations, TE Connectivity is telling its end-to-end optical story, and Andy Bechtolsheim is back with a take on pluggable optics that you are going to want to hear. Let's get into it. Skyler: Alright, first up — there's a new organization in town focused on optical interconnects for A I datacenters, and it's worth paying attention to who's at the table. A group of leading optical connectivity solutions providers just announced the formation of a new industry body specifically aimed at supporting optical interconnects for A I datacenter applications. Now, we don't have the full membership list yet, but when vendors in the optical space decide to get organized, it usually means a few things are happening at once — the market is real enough to warrant standards alignment, there's enough divergence in approaches that interoperability is becoming a problem, and somebody needs to define what "good" looks like before the hyperscalers do it for them unilaterally. Skyler: So why does this matter right now? Because A I training clusters are fundamentally different from traditional datacenter workloads when it comes to optical requirements. You're not just moving packets — you're sustaining massive, sustained bandwidth between thousands of accelerators, and the latency and loss characteristics matter in ways they never did for web traffic. The optical interconnect layer is under enormous pressure. You've got co-packaged optics emerging as a potential long-term answer, pluggable modules being pushed to their limits at 800G and 1.6T, and now active electrical cables and linear drive pluggables trying to carve out niches in between. Skyler: When you have that many competing form factors and approaches, you need some kind of industry alignment — otherwise you end up with a situation where every hyperscaler is doing something custom and the ecosystem fragments. The last thing a Tier 2 cloud provider or an enterprise building an A I cluster needs is to pick the wrong optical form factor and find themselves stranded. That's exactly the kind of problem a well-structured industry consortium can help prevent — common specs, defined interoperability test suites, shared reference architectures. Skyler: The key question I have — and we'll be watching this — is whether this new org has teeth. Some consortiums produce genuinely useful specs that become defacto standards. Others produce white papers that nobody reads. The differentiator is usually whether the hyperscalers and the silicon vendors get involved, because they're the ones whose purchasing decisions actually shape the market. If Google, Microsoft, and Meta are in the room, or at least aligned with what's coming out of this group, it matters. If it's purely vendor-driven, it's going to have a harder time gaining traction. Worth keeping an eye on who joins over the next few months. Skyler: Okay, story two — Synopsys just announced silicon IP cores for both Ultra Ethernet and U A Link, and this is genuinely significant for the state of the A I fabric market. Let me explain why. For a while, the criticism of both U E C — the Ultra Ethernet Consortium's spec — and U A Link was that they were consortium initiatives that existed mostly on paper. Great ideas, important work, but where's the actual implementable silicon? Well, Synopsys just answered that question, at least partially. Skyler: For those less familiar — Ultra Ethernet, or U E C, is an effort to extend Ethernet specifically for A I workloads. Standard Ethernet wasn't designed for the all-to-all collective communication patterns you see in large model training. Things like RDMA over Converged Ethernet — RoCE — helped, but U E C is going further, rethinking transport-level behavior for A I traffic, including stuff like out-of-order handling, congestion response, and collective operations support. It's Ethernet, but rebuilt for the job. Skyler: U A Link is a different beast — it's the scale-up fabric, the thing that connects accelerators within a node or within a tightly coupled pod. Think of it as the industry's answer to NVLink — an open standard for that ultra-low-latency, high-bandwidth coherent fabric that A I accelerators need when they're working in tight coordination. The fact that Synopsys is now offering verified, licensable IP for both of these is a real maturation signal. Skyler: Here's why it matters practically: if you're a chip company — an A I accelerator startup, or even a larger player building a custom SoC — you no longer have to build your U E C or U A Link PHY and controller logic from scratch. You license the IP from Synopsys, you integrate it, you tape out. That dramatically reduces the time and risk associated with adopting these standards. It's the same reason ARM architecture became dominant — silicon IP availability is a massive accelerant for ecosystem adoption. Skyler: Now, some caveats. U E C spec is still evolving. We're not at a point where everything is locked down and ratified end to end. So there's a version-compatibility question — IP licensed today will need to track the spec as it finalizes. And U A Link is even earlier in its journey in terms of broad adoption. But Synopsys shipping this IP is a signal that the standards have enough stability that a major EDA and IP house is willing to make a commercial bet on them. That's meaningful. Watch for other IP vendors to follow, because Synopsys doesn't usually play in a market alone for long. Skyler: Moving on — Cisco dropped a blog post about something they're calling Nexus One Fabric, and it's worth decoding what they're actually saying here. On the surface, it's a pitch for using open VXLAN and EVPN standards to unify datacenter fabric operations. But the subtext is more interesting — Cisco is positioning this as an answer to the sprawl problem that comes when you have multiple separate network overlays for different workloads, and A I is making that problem much worse. Skyler: Here's the context. Traditional enterprise datacenters were already running VXLAN and EVPN for server virtualization, storage, and general compute. Then A I training clusters came in — and they often run on completely separate fabrics. You've got your RoCE fabric for the GPU-to-GPU traffic, your front-end Ethernet for storage and management, maybe an out-of-band network, and suddenly the ops team is managing three or four conceptually different networks. That's expensive and complex. Skyler: What Cisco is arguing with Nexus One Fabric is that you can collapse some of that complexity by standardizing on VXLAN and EVPN as a common control and data plane across workload types. EVPN gives you a scalable, standards-based control plane for distributing MAC and IP reachability information, and VXLAN gives you the overlay encapsulation to carry traffic across an underlay without caring about the physical topology. Both are genuinely well-understood, widely deployed standards — this isn't vaporware. Skyler: The interesting tension here is that the A I infrastructure crowd often looks at standard Ethernet-based overlays with some skepticism for high-performance training workloads. VXLAN adds encapsulation overhead, and when you're trying to squeeze every nanosecond out of GPU-to-GPU communication, that matters. Cisco knows this, and their answer is essentially: yeah, but for most of the traffic in your datacenter — management, storage access, inter-cluster communication — unified operations with open standards beats running bespoke fabrics everywhere. They're not saying use VXLAN for your innermost training collective comms. They're saying stop running five different network management systems. Skyler: It's a pragmatic pitch, and honestly for operators running mixed A I and traditional enterprise workloads, there's something to it. The question is whether the hyperscalers — who have already built their own custom underlay fabrics — care. This is probably more relevant to the Tier 2 cloud market, the enterprises building serious A I capability, and the colocation providers trying to offer A I-ready infrastructure without a custom silicon team. Skyler: Story four — TE Connectivity published a piece in Fibre Systems about scaling end-to-end optical interconnects for A I datacenters, and I want to talk about this because TE is approaching the problem from an angle that doesn't get enough coverage. Most of the conversation about optical interconnects centers on the transceiver — the pluggable module, or whether co-packaged optics will replace it. TE is coming at it from the physical layer infrastructure side — connectors, cables, backplanes, and system-level integration — and that stuff is genuinely hard and genuinely underappreciated. Skyler: Here's the challenge they're addressing. As you scale optical interconnects across a datacenter — from within a rack, across a row, across a datacenter floor — you're not just dealing with the optics themselves. You're dealing with routing density, bend radius limitations of fiber, connector loss budgets, thermal management around high-power optical components, and the mechanical reliability of connections that will be inserted and removed thousands of times over a system's lifetime. When you're at 400G per lane and heading to 800G, every decibel of insertion loss matters. A connector that added negligible loss at 10G is suddenly a real budget item. Skyler: TE's thesis is that you have to think about this as a system problem, not a component problem. You can have the best transceiver in the world, but if your MPO connectors are dirty, your cable management is causing excessive bend loss, or your backplane is introducing reflections, your link budget falls apart. And at the scale of an A I cluster — we're talking tens of thousands of optical connections per pod — even a small per-connection inefficiency multiplies into a massive problem. Skyler: What's interesting about TE's position is that they're essentially arguing for vertical integration of the physical optical infrastructure. Not the optics themselves, but the system that hosts and connects the optics. Cassette systems, structured cabling optimized for high-density A I rack layouts, connectors designed for the specific loss budgets of co-packaged and near-package optical architectures. It's less glamorous than a new transceiver announcement, but it's the kind of thing that separates a datacenter that actually hits its optical performance targets from one that's constantly chasing link errors. Skyler: The broader point for the industry is that the move toward higher-density, higher-bandwidth optical connectivity is forcing a rethink of the entire physical infrastructure stack. It's not enough to spec out the ASIC and the transceiver — you have to spec out the fiber plant, the connector system, and the cable routing architecture from the start. Companies like TE are betting that datacenter builders will start treating this as a first-class design constraint rather than an afterthought, and given the capex at stake in modern A I cluster builds, that bet looks pretty solid. Skyler: And now for story five, which is the one I've been looking forward to — Andy Bechtolsheim and a group of co-founders are apparently not done with pluggable optics yet. The Next Platform ran a piece headlined "Bechtolsheim and Friends Breathe Life Into Pluggable Optics One Last Time," and that framing alone tells you something about the stakes and the drama in this space. Skyler: For those who don't know the name — Andy Bechtolsheim co-founded Sun Microsystems, was the first investor in Google, co-founded Arista Networks, and has basically been at the center of datacenter networking hardware for four decades. When he bets on something, people pay attention. And right now, he's apparently betting that pluggable optics have more runway than the co-packaged optics narrative would suggest. Skyler: Here's the tension this story is sitting inside. The dominant narrative for the last couple of years has been that co-packaged optics — where the optical engine is integrated directly onto the switch package, much closer to the switching silicon — is inevitably going to displace pluggable optics for high-bandwidth datacenter applications. The argument is fundamentally about power and bandwidth density. At 1.6T and beyond per port, running electrical signals from a switch chip across a PCB to an external pluggable module wastes power and is getting harder to do reliably. Co-packaged optics solves the electrical distance problem by moving the optical conversion right up against the die. Skyler: But pluggable optics have advantages that are hard to give up. They're field-replaceable. They're hot-swappable. They support mix-and-match from multiple vendors. The ecosystem is mature, the supply chain is deep, and operators know how to manage them. Co-packaged optics, by contrast, means if your optical engine fails, you may be looking at returning the whole switch. That's a very different operational model, and for a lot of datacenter operators, it's a scary one. Skyler: What Bechtolsheim and his collaborators seem to be working on — and the details are somewhat sparse at this stage — is a new take on pluggable optics that pushes the performance envelope enough to stay competitive with co-packaged at the next generation of bandwidth. Think linear drive architectures, pushing the DSP complexity out of the module to reduce power, or higher-density form factors that change the bandwidth-per-watt math. The "one last time" framing in the headline is a bit dramatic — it implies this is the last stand for pluggable. But I'd push back on that slightly. Skyler: History suggests that every time co-packaged optics seems inevitable, pluggable finds another gear. The ecosystem around pluggable is enormous — the component vendors, the module makers, the testing infrastructure — and that kind of ecosystem inertia is real. What I think Bechtolsheim is doing is making a bet that pluggable can get to 1.6T and maybe 3.2T in ways that are power-competitive enough that operators will choose the operational familiarity of pluggable over the theoretical density advantages of co-packaged. Whether he's right is one of the most interesting open questions in datacenter optics right now. But I would not bet against him. Skyler: Alright, that's a wrap on today's episode. We covered a new optical interconnect consortium finding its footing, Synopsys making U E C and U A Link real with actual licensable IP, Cisco's push for unified fabric operations with open standards, TE Connectivity's system-level case for rethinking physical optical infrastructure, and Andy Bechtolsheim's bet that pluggable optics still has a chapter left to write. Huge thanks to our sponsors — Ada, who helps organizations actually integrate A I into their real-world workflows, and Zen Semiconductor, an umbrella company with a portfolio across the silicon industry, including Ada. This is Line Rate, brought to you by Ada and Zen Semiconductor. If you found this useful, subscribe wherever you get your podcasts, share it with someone building or investing in A I infrastructure, and check the show notes for links to everything we covered today. See you next time.