Co-Packaged Optics Market (2035): Optical Engines, Silicon - openPR.com

Skyler: Welcome to Line Rate — the show about datacenter fabric, A I interconnect, and the hardware that's reshaping how the world's biggest compute clusters move data. I'm Skyler Monroe, and if you care about what's happening at the physical layer of A I infrastructure, you're in the right place. Huge thanks to our sponsors: Ada, who helps enterprises actually integrate A I into their workflows in meaningful ways, and Zen Semiconductor, an umbrella company with a portfolio of businesses spanning the silicon industry, including Ada on the deployment side. Today we've got a packed show — co-packaged optics market projections, Meta's open networking vision from OCP Summit, Marvell's take on multi-protocol scale-up interconnect, a significant investment into Singapore-based DenseLight, and a fully liquid-cooled A I switch that's turning heads. Let's get into it. Skyler: Alright, first up — the co-packaged optics market and where it's headed by 2035. Now, market research reports can be hit or miss, but this one is worth talking through because the underlying dynamics it's tracking are very real, and they're reshaping decisions being made right now in silicon and systems design. The headline is that CPO — co-packaged optics — is moving from a niche, bleeding-edge experiment to what looks like a central pillar of A I datacenter networking architecture over the next decade. Skyler: So let's step back for a second. Why does CPO matter? Traditional pluggable optics — your QSFP-DD, your OSFP modules — they sit at the edge of your switch ASIC, connected by traces on a PCB. Every hop across that PCB burns power and introduces signal degradation. As we push to 800G per lane and beyond, that loss budget gets tighter and tighter. Co-packaged optics essentially collapses that distance — you're integrating the optical engine right next to, or in some cases on top of, the switch ASIC. The electrical distance drops dramatically, which means you spend less power on SerDes, you get better signal integrity, and in theory you can push to higher aggregate bandwidths. Skyler: The market report focuses on the optical engines themselves — the laser, modulator, and detector components — as the key value-add layer. And that makes sense. The silicon photonics platforms and the III-V laser integration approaches are where the real differentiation is happening. You've got companies betting on external cavity lasers that couple into silicon waveguides, and others going for monolithic integration. Neither approach is clearly winning yet, and that's actually part of what makes this market so interesting to watch. Skyler: The 2035 horizon feels far away, but think about where we are in the adoption curve. We're still in the early commercial phase — a handful of hyperscalers are doing proof-of-concept deployments, Broadcom and Cisco have made public commitments to CPO roadmaps, and the ecosystem of optical engine suppliers is slowly consolidating. By the early 2030s, if the cost curves play out the way proponents expect, CPO could become the default architecture for high-radix spine switches in large A I clusters. That's a massive shift from where we are today. Skyler: For engineers in the audience — the key technical challenges that will determine whether this timeline holds are: thermal management of the optical engine at package level, the yield and reliability of the photonic integrated circuits, and standardization of the optical engine interface so switch vendors aren't locked into a single optics supplier. That last one is politically as much as technically complicated. Watch the co-packaged optics working groups in OIF and IEEE closely — they're where the interface specs are being hammered out, and progress has been slower than some hoped. Skyler: Bottom line: CPO is real, the trajectory is real, 2035 as a maturation point is plausible, and the companies that solve the thermal and yield problems first are going to have enormous leverage in next-generation A I switch design. Skyler: Next, let's talk about OCP Summit 2025 and what Meta's engineering team put out about the open future of networking hardware for A I. This is a meaty one because Meta is not just talking strategy — they're laying out an architectural philosophy that they're actively building to, and it has real implications for the broader ecosystem. Skyler: So the core thesis from Meta's team is essentially: proprietary networking stacks are a liability at hyperscaler scale. When you're running clusters in the hundreds of thousands of GPUs, vendor lock-in on your network fabric isn't just a procurement headache — it's an engineering constraint. You can't iterate fast enough, you can't customize deeply enough, and you end up co-dependent on roadmaps that don't perfectly align with your training and inference workload evolution. Skyler: The open networking angle here is interesting because it's not just about open-source software — it's about open hardware specifications. Meta has been contributing to OCP for years, but what's notable now is the level of specificity around A I fabric requirements. They're talking about low-latency collective operations, congestion control behavior tuned for all-reduce workloads, and telemetry architectures that give you real-time visibility into fabric utilization at a granularity that traditional data center monitoring was never designed to provide. Skyler: One thing I want to highlight: Meta is explicitly calling out the importance of open optical interfaces in the context of their scale-out fabric. This ties directly back to the CPO story — if you're going to have an open, disaggregated network where you're mixing and matching components from different vendors, you need standardized optical interfaces at every layer. That's a strong signal for the OIF work I mentioned earlier, and it's also a signal to optical engine startups that hyperscaler demand for open interfaces is real and growing. Skyler: The competitive read here is interesting. Meta being vocal about open networking at OCP creates a kind of permission structure for other hyperscalers to push back on proprietary switch vendors. It validates the merchant silicon approach — when Meta says "we want to build around open ASICs and open optical interfaces," that's a green light for Broadcom, Marvell, and others to compete aggressively on the silicon layer without worrying that the hyperscalers are going to retreat to full vertical integration. Skyler: But — and this is an important caveat — open doesn't mean undifferentiated. Meta is still going to build custom software, custom management planes, custom congestion control algorithms. The openness is at the hardware interface layer. So for companies building A I networking gear: the opportunity is to win on hardware performance and programmability, not on locking customers into software ecosystems. That's a different business model than traditional networking, and not every incumbent is comfortable with it. Skyler: There's also a talent and ecosystem angle here. When Meta publishes detailed specs at OCP, it pulls in the broader engineering community — startups, academics, second-tier cloud providers — and accelerates the ecosystem around open A I fabric. That's a strategic advantage for Meta even if they're giving away information, because they shape the direction of the ecosystem to their benefit. Skyler: Okay, let's pivot to Marvell, who put out a piece specifically on scale-up networking for A I, focusing on multi-protocol interconnect and optics. This is Marvell in full roadmap-communication mode, and there's a lot of substance to unpack. Skyler: Scale-up, just to be precise for anyone who's newer to the terminology — scale-up refers to the high-bandwidth, ultra-low-latency interconnect within a single A I training job, typically within a rack or across a small number of racks. This is where your NVLink-class and U A Link-class connectivity lives. Scale-out is the broader fabric that connects racks and pods. The reason this distinction matters is that the performance requirements are very different — scale-up needs single-digit microsecond latency and very high bandwidth density, while scale-out can tolerate slightly higher latency in exchange for massive aggregate bandwidth and routing flexibility. Skyler: Marvell's angle is multi-protocol, and this is strategically savvy. Rather than betting that one interconnect protocol wins — say, PCIe Gen6, or CXL, or U A Link, or NVLink — they're building silicon and optics solutions that can support multiple protocols over a common physical layer. That's a hedge, but it's also a real architectural advantage. If you're a hyperscaler or a Tier-2 cloud trying to build a heterogeneous accelerator cluster with GPUs from different vendors, you need a fabric layer that doesn't force protocol homogeneity. Skyler: The optics piece of Marvell's announcement is directly tied to extending scale-up distances. Traditionally, scale-up interconnects are copper or active electrical cables over short distances — a few meters at most. But as GPU clusters get larger and the optimal physical topology demands longer reach, you need optics in the scale-up fabric. Marvell is positioning their optical DSP and silicon photonics IP as the bridge between scale-up protocols and optical transport, which is a non-trivial integration challenge. Skyler: I want to flag the U A Link angle specifically. U A Link is the consortium-backed open alternative to NVLink for scale-up GPU interconnect, and its specs are still actively evolving. Marvell has been involved in the U A Link consortium, and their multi-protocol messaging is partly a signal to U A Link members that they can deliver silicon that meets the U A Link physical layer requirements. But U A Link is not finalized — the spec is still being worked, adoption timelines are uncertain, and NVLink remains dominant in NVIDIA-centric deployments. So when Marvell says multi-protocol, read that as a bet that the market will fragment by protocol across different accelerator vendors, and they want to be the neutral silicon provider to all of them. Skyler: The competitive picture: Marvell is competing here with Broadcom, whose Bailly and Humboldt series targets similar scale-out and scale-up markets, and increasingly with startup silicon vendors who are building purpose-built scale-up fabrics. Marvell's advantage is breadth — they have optical DSP, PAM4 retimer, and switching ASIC IP all in house, which lets them offer a more integrated solution than a pure-play startup. Their challenge is that breadth can also mean you're not the absolute best at any single thing. For customers who need maximum performance on a specific protocol, purpose-built silicon is often more appealing. Skyler: All right, story four — and this one is a funding story that I think is genuinely underappreciated: Lite-On Technology is investing 34.3 million dollars in DenseLight Semiconductors, a Singapore-based optical firm, specifically targeting A I datacenter optical interconnects. Let me tell you why this matters beyond the headline number. Skyler: DenseLight is not a household name in the way that II-VI or Coherent or Lumentum are, but they have been doing serious work in III-V photonics — indium phosphide specifically — for a long time. InP is critical for high-speed lasers and modulators, and it's one of the materials where co-packaged optics architectures need to make a choice: do you couple external InP lasers into silicon photonic waveguides, or do you try to integrate III-V gain material directly on the silicon platform? DenseLight's expertise sits right at that intersection. Skyler: Lite-On investing 34.3 million dollars here is significant for a few reasons. Lite-On is a major Taiwanese ODM and components manufacturer — they make power supplies, storage, and optical components for major OEMs and hyperscalers. They're not a speculative investor; when Lite-On puts money into an optical firm, it's because they see a clear path to integrating that technology into products they can sell at volume to their customer base. That customer base includes some of the largest server and networking OEMs in the world. Skyler: The Singapore angle is also interesting. DenseLight is based in Singapore, which has a growing semiconductor and photonics ecosystem with government support behind it. Singapore is trying to build out III-V photonics manufacturing capability, and a 34.3 million dollar investment from a major Taiwanese strategic partner is a meaningful vote of confidence in that ecosystem. From a geopolitical supply chain perspective, having InP laser capability in Southeast Asia is increasingly valuable as the industry tries to diversify away from concentration risk. Skyler: For the A I datacenter angle specifically — the investment is explicitly called out as targeting optical interconnects for A I data centers. That means Lite-On sees DenseLight's technology fitting into the co-packaged optics supply chain, likely as a supplier of the laser and optical gain elements that get integrated into CPO modules. If that's right, DenseLight could be supplying into the same switches we talked about in the CPO market story — the scale-up and scale-out fabric for next-generation A I clusters. Skyler: The broader pattern here is consolidation and vertical integration in the photonics supply chain. As CPO moves toward commercial scale, the optical engine vendors need reliable, high-performance laser sources, and the laser source companies need capital to scale manufacturing. Strategic investments like this — where the investor has a clear route to market — are exactly how that supply chain consolidation happens. Watch for more of these deals as we get into 2025 and 2026. Skyler: Last story, and this one is fun — a new A I switch that is fully liquid-cooled and designed to reduce power consumption. This is the kind of product announcement that sounds incremental until you think about what it actually implies for datacenter design. Skyler: So the headline is a fully liquid-cooled A I switch. Most network switches today are air-cooled — you've got front-to-back or side-to-side airflow, fans, the whole arrangement. A small number of high-density switches have started appearing with some liquid cooling assist, but a fully liquid-cooled switch is still relatively unusual. The reason this is interesting right now is the thermal context: switch ASICs are getting dramatically more power-hungry as radix and bandwidth scale. A 51.2 terabit-per-second switch ASIC is consuming north of 300 watts just for the ASIC itself. Add in the optics and you're well over 500 watts for the forwarding plane alone. Air cooling that in a dense rack is genuinely hard. Skyler: Liquid cooling solves the heat removal problem much more efficiently — water has about 4,000 times the heat capacity of air by volume, which means you can extract the same heat with dramatically less infrastructure and noise. For A I clusters specifically, where you're trying to pack as many GPUs and as much networking as possible into a given footprint, liquid-cooled switches mean you can deploy higher-density networking without running into thermal limits or blowing your facility's HVAC budget. Skyler: The "designed to use less power" framing is important to parse carefully though. Liquid cooling doesn't inherently make the ASIC consume less power — the silicon still burns what it burns. What changes is the efficiency of heat removal and potentially the ability to operate at higher performance states without thermal throttling. So the power story is more nuanced: the switch itself might draw similar power, but the facility-level power consumption for cooling infrastructure can drop significantly when you move from chilled air to direct liquid cooling. That's a real OpEx advantage for hyperscalers and large cloud providers who are paying massive electricity bills. Skyler: There's also a co-packaged optics angle here — and yes, CPO keeps coming up because it's genuinely central to where this is all going. If you're integrating optics into the switch package, you now have another heat source inside or adjacent to the switch ASIC, and that makes the thermal management problem significantly harder for air-cooled designs. A liquid-cooled switch chassis is much better positioned to handle CPO thermal loads, which means the companies building liquid-cooled switch platforms today may have a meaningful advantage when CPO becomes mainstream. Skyler: The competitive landscape for A I switches is getting very interesting. You've got Arista, Cisco, and Juniper on the traditional networking side, Nvidia with their Quantum and Spectrum platforms pushing hard into A I fabric, and a growing set of challengers — both established players like Marvell and Broadcom providing ASICs to ODMs, and startups building purpose-built A I network hardware. A liquid-cooled switch that can operate at higher thermal envelopes is a differentiator today, and it's likely to look prescient in two or three years when the next generation of higher-power ASICs arrives. Skyler: And honestly, the liquid cooling angle is also a facility design signal. When networking gear starts showing up with liquid cooling as a first-class feature, it means datacenter operators need to plan for liquid delivery to the network layer, not just to the compute layer. That's a non-trivial change to facility design and to the conversations between network architects and facility teams. Get those conversations started early. Skyler: Alright, that's a wrap on today's Line Rate. We covered a lot of ground — CPO market dynamics through 2035, Meta's open networking vision from OCP Summit, Marvell's multi-protocol scale-up interconnect push, Lite-On's strategic bet on DenseLight in Singapore, and a fully liquid-cooled A I switch that's pointing toward where datacenter thermal management is headed. This is a space where a lot is moving fast and a lot is still half-baked, and that's exactly why it's worth paying close attention. Line Rate, brought to you by Ada — helping organizations integrate A I into their workflows — and Zen Semiconductor. If you found this valuable, please 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 discussed today. I'm Skyler Monroe — see you next episode.