Sivers Semiconductors, O-Net and Enablence Technologies Announce Collaboration on Co-Packaged Optics - TMX Newsfile

Skyler: Hey, welcome back to Line Rate — the show that tracks everything happening at the bleeding edge of datacenter fabric and AI high-speed interconnect. I'm Skyler Monroe, and if it moves bits between chips, racks, or datacenters at ludicrous speed, we're talking about it here. Big thanks to our sponsors: Ada, who helps companies actually integrate A I into their real workflows, and Zen Semiconductor, an umbrella company with a portfolio of businesses across the silicon industry — and they work with Ada on the A I deployment side too. Alright, we've got a stacked show today — co-packaged optics is heating up with a new three-way collaboration, coherent optics and fiber are getting their investor moment, Cisco is making the case for Ultra Ethernet in A I clusters, Ciena is pitching something called hyper-rail photonics, and we're going to dig into what system-level 1.6 terabit-per-second interoperability actually demands. Let's get into it. Skyler: Okay, first up — Sivers Semiconductors, O-Net, and Enablence Technologies have announced a collaboration on co-packaged optics. Three companies, one announcement, and it's worth unpacking why each of them is in the room. Sivers is a Swedish company that does RF and photonics semiconductors — they've been pushing hard into silicon photonics and laser components. O-Net is a Chinese optical module and transceiver manufacturer with real scale. And Enablence is a photonic integrated circuit company that's been around the waveguide and planar lightwave circuit space for years. So what you have here is a laser source, a photonic integration play, and a volume manufacturer all deciding to work together on CPO. And that combination is not an accident. Skyler: Co-packaged optics, if you're coming at this fresh, is the approach where instead of plugging a separate optical transceiver into a switch ASIC — the way we do it with QSFP-DD or OSFP pluggables today — you integrate the optical engines much closer to the switch silicon itself. Right on the package, or extremely close to it. The reason this matters is power and bandwidth density. At 51.2 terabits per second on a single switch ASIC, which is where Tomahawk 5 and similar products are today, driving all those SerDes lanes across a board to pluggable cages and then back is burning a huge amount of power in the electrical domain. CPO moves the optical conversion point much closer to the die and can dramatically cut that electrical overhead. Skyler: The challenge — and it's been a stubborn one — is that CPO requires extremely tight integration across laser sources, photonic integrated circuits, and the switch ASIC itself. The thermal environment is brutal. The coupling tolerances are tight. And you need all of those pieces to come from somewhere, qualify together, and be manufacturable at volume. That is exactly what this three-way partnership is trying to address. Sivers brings the laser and photonic semiconductor expertise. Enablence brings waveguide integration and PIC design. O-Net brings manufacturing scale and optical assembly experience. It's a vertically coordinated supply chain in embryonic form. Skyler: Now, is this a shipping product? No. This is a collaboration announcement, and in CPO land, there's a well-established gap between announcing and delivering. We've seen CPO demos and partnerships dating back to the early 2020s and the technology is still not mainstream in production switches. But the cluster of companies trying to solve this problem keeps growing, and that's the real signal. When you see a Swedish photonics semiconductor company, a Chinese optical module house, and a Canadian PIC company deciding to align, it tells you the ecosystem is starting to coalesce around what a CPO supply chain actually needs to look like. This one's worth watching. Skyler: Next story — there's a piece making the rounds from Dr. Robert Castellano's Substack arguing that Coherent and Corning are more strategically important to the A I buildout than most investors realize. And honestly, this is a framing I find pretty compelling, even if the audience for that piece is more financially oriented than technically oriented. Let me give you the version of this argument that actually makes sense for people who live in the stack. Skyler: The thesis is simple: the conversation about A I infrastructure spending is dominated by GPUs, HBM, and the big switch ASICs. But none of that compute scales without the optical layer holding it together. And right now, that optical layer is being pushed extremely hard. Coherent — formerly II-VI, formerly Finisar, it's had a lot of names — is one of the very few companies with the vertical integration to produce transceivers from laser epitaxy all the way through to shipping optical modules. That's a rare capability. They're in coherent DWDM for the long-haul and metro links, but they're also pushing into the 800G and 1.6T pluggable space that's directly relevant to front-end networks in large A I clusters. Skyler: Corning's angle is different. They make the fiber. And this sounds boring until you remember that the kind of fiber you use matters a lot when you're running 800G or 1.6T signals over meaningful distances inside or between datacenters. Corning's ultra-low-loss single-mode fiber and their multimode fiber for shorter reach are both in very high demand, and the supply of fiber isn't infinitely elastic. It takes time to build fiber draw capacity. So the argument is: even if you fully believe in the GPU and ASIC spending cycle, the optical transport layer — transceivers and fiber — is a non-negotiable dependency, and it's capacity-constrained in ways that the market may not be fully pricing. Skyler: From a technical standpoint, what I'd add is that coherent optics is increasingly showing up inside the datacenter, not just between datacenters. There are architectures being seriously evaluated where coherent DSP-based transceivers handle the longer campus-scale or inter-building runs that used to be done with direct detect or active optical cables. That extends Coherent's addressable market significantly. And Corning benefits from any architecture that uses more single-mode fiber — CPO transitions and coherent inside the datacenter both point in that direction. So whether you're an investor or an architect, the optical layer deserves more attention than it usually gets in the A I infrastructure conversation. Skyler: Alright, story three — Cisco published a blog post making the case for Ultra Ethernet in scalable A I network deployments. This is Cisco putting their flag in the ground on UEC, and it's worth reading carefully because Cisco's position here reflects something real about where the market is going. Skyler: Quick context: the Ultra Ethernet Consortium — UEC — is an industry group that formed in 2023 to define Ethernet-based transport enhancements specifically for A I and HPC workloads. The core problem they're trying to solve is that standard Ethernet, even at high speeds, has behaviors that are really unfriendly to the all-to-all communication patterns of distributed A I training. Things like head-of-line blocking, the way congestion spreads, and latency variability that can cause massive stragglers in a training job. UEC is defining new transport-layer behaviors — things like out-of-order packet reassembly, better multipath load balancing, improved congestion notification — all while staying on Ethernet as the foundation. The idea is you get the scale, the ecosystem, and the cost structure of Ethernet, but you fix the parts that make it painful for A I. Skyler: Cisco's blog is framing this as the path to scalable A I networks that don't require going to InfiniBand. And that's the competitive lens here — InfiniBand, specifically NVIDIA's NVLink and Quantum InfiniBand stack, currently dominates the backend fabric for large A I training clusters. The case against InfiniBand is: it's expensive, it's largely a single-vendor ecosystem, and at some point the industry wants an open, Ethernet-native alternative. UEC is attempting to be that alternative. Skyler: What Cisco is specifically highlighting is the combination of UEC's transport layer improvements with high-radix Ethernet switching at 51.2T or eventually 102.4T switch ASICs, and how that combination can support the cluster sizes and traffic patterns that modern A I training demands. They're also noting the role of RDMA over Ethernet — RoCE — and how UEC complements that. Now, I want to be clear about where UEC stands: the spec is still maturing. The 1.0 specification has been published but implementations are early. There are open questions about how UEC behavior interacts with existing RoCE deployments, and about the end-to-end latency numbers in real mixed-traffic scenarios. Cisco's blog is advocacy, and it's well-reasoned advocacy, but UEC is not a finished, widely-deployed thing yet. The next twelve to eighteen months will be telling. Skyler: Story four — Ciena is out making the case for what they're calling hyper-rail photonics in datacenter fiber networks. This one is interesting because Ciena is primarily a carrier and metro optical networking company — they're known for WaveLogic coherent DSPs and high-capacity DWDM systems. So when they start talking about datacenter fiber architectures, it means they see an opportunity to extend their technology and their market. Skyler: The hyper-rail concept — and this is a term with some variation in how different companies use it — is essentially about creating dedicated, high-capacity optical paths within a datacenter that serve A I workloads specifically. The idea is that instead of treating all datacenter traffic the same way over a common switched fabric, you build photonic rail infrastructure that provides direct, low-latency, high-bandwidth optical connectivity between specific endpoints — between GPU servers or between racks — without going through multiple layers of electrical switching. Think of it as optical bypass or optical circuit switching as a complement to the packet-switched layer. Skyler: Ciena's specific angle is bringing coherent optical technology — the kind of technology that carries terabits across metro and long-haul — into the datacenter to handle these high-bandwidth point-to-point or point-to-multipoint flows. The argument is that for the traffic patterns of A I training, where you have very large, sustained flows between specific pairs or groups of accelerators, a coherent optical rail can be far more efficient than hairpinning everything through a packet switch. You get massive bandwidth, very low latency, and you remove hops. Skyler: The practical question is always integration and cost. Coherent DSP chips have traditionally been expensive and power-hungry relative to what's used inside a datacenter. Ciena and others are working hard to bring those costs down, and the latest generation of WaveLogic ASICs is dramatically more efficient than what came before. But you're still adding infrastructure — you're adding an optical layer with its own management plane, its own provisioning, its own failure modes. For hyperscalers building at scale, that might absolutely be worth it. For an enterprise building a smaller A I cluster, the tradeoff calculus looks different. Ciena is clearly targeting the hyperscaler and large-scale A I infrastructure segment, and that's where this technology has the most compelling near-term case. Skyler: And story five — Semiconductor Engineering ran a piece on system-level design for 1.6 terabit-per-second interoperability in A I datacenters. This is a technical deep-dive, and it's exactly the kind of content that matters a lot more than people outside the hardware industry realize. Skyler: Let me set the scene. 1.6 terabits per second per port is the next major milestone in Ethernet and optical interconnect for A I datacenters. You're looking at either 2x800G or 8x200G or various lane configurations depending on the PHY approach. Getting to 1.6T is not just about turning up the speed — every layer of the stack has to co-evolve. The SerDes lanes running at 112 gigabaud or pushing toward 224 gigabaud. The electrical package traces. The optical engine, whether that's a pluggable module or a co-packaged solution. The connectors and the PCB materials. All of it has to work together, and more importantly, it has to interoperate across vendors. Skyler: That last point — interoperability — is where the Semiconductor Engineering piece focuses, and it's where things get genuinely hard. At 800G, the industry went through a fairly painful process of getting different vendors' optical modules and switch ASICs to reliably interoperate. At 1.6T, the tolerance budgets get tighter, the signal integrity challenges increase, and you have new variables like whether you're using linear pluggable optics — LPO — or traditional DSP-based modules. LPO is an approach where you eliminate or minimize the DSP in the optical module and lean on the host-side DSP in the switch ASIC to do more of the equalization work. That can save power and cost in the module, but it pushes more burden onto the system-level design and makes interoperability harder because you're coupling the behavior of the host and the module much more tightly. Skyler: The piece walks through how system-level design — meaning how you co-design the ASIC, the package, the board, and the optical interface together — becomes non-negotiable at 1.6T. You can't just buy a switch ASIC, plug in a module from vendor B, and assume it works. You need to understand the channel loss budgets, the equalization headroom, the FEC overhead, and how all of those interact. Multi-source agreements — MSAs — try to define the interface specs so that interoperability is possible, but the margins at 1.6T are thin enough that "in spec" doesn't always mean "works great in practice." The industry is actively working on this, and bodies like OIF and IEEE are involved in nailing down the specs. But anyone who's planning a 1.6T deployment in the next eighteen months should be thinking hard about system-level validation, not just component-level spec compliance. Skyler: Alright, that's a wrap on today's show. Co-packaged optics is finding its ecosystem, the optical layer is getting the investor attention it deserves, Ultra Ethernet is making its case against InfiniBand, hyper-rail photonics is bringing coherent into the datacenter, and 1.6T interoperability is harder than the spec sheets suggest. Big week for the fabric world. Line Rate, brought to you by Ada and Zen Semiconductor — Ada helps companies integrate A I into their real-world workflows, and Zen Semiconductor is an umbrella company with a portfolio across the silicon industry. If you got value out of this episode, please subscribe and share it with someone who's building or buying A I infrastructure — this stuff moves fast and the more people who understand it the better. Show notes and links are at the usual place. I'm Skyler Monroe, catch you next time on Line Rate.