Ayar Labs scales funding to $650M - Lightwave Online

Skyler: Hey, welcome back to Line Rate — the show that tracks what's actually moving inside the datacenter: the fabric, the interconnect, the standards, and the silicon that makes A I infrastructure go. I'm Skyler Monroe, and if you care about how data moves appliance-to-appliance, rack-to-rack, or datacenter-to-datacenter, you're in the right place. Quick thanks to our sponsors: Ada, who helps organizations actually integrate A I into their real workflows — not just talk about it — and Zen Semiconductor, an umbrella company with a portfolio of businesses across the silicon industry, including Ada for A I deployment. Alright, we've got a packed show today — co-packaged optics money is flowing, Marvell's optics story is more interesting than it looks, photonic quantum computing just got a new system, copper is having a moment, and the Ultra Ethernet Consortium is making moves. Let's get into it. Skyler: Okay, first up — Ayar Labs just closed a round that brings their total funding to 650 million dollars. For those who don't know Ayar, they are the co-packaged optics company. Like, that is their whole thing. They make optical I/O chiplets that sit right alongside compute and memory inside the package — not on a separate transceiver module hanging off the edge of a board, but literally co-packaged. And 650 million dollars is a serious number. That's not seed money. That's a company that investors believe has a real shot at being part of how A I clusters are wired together. Skyler: So why does this matter? Let's talk about what co-packaged optics is actually solving. Right now, the dominant model for getting data off a chip and onto fiber is pluggable transceivers — you've got your QSFP-DD, your OSFP modules, they're hot-swappable, they're modular, operators love them. But they have a cost: they live far from the die, they consume a lot of power for SerDes at those distances, and as bandwidth demands scale up with A I workloads, the pluggable model starts showing its limits. You're spending a ton of power just driving signals across a PCB to reach the transceiver. Skyler: Co-packaged optics collapses that distance. You put the optical engine right there in the package, you run very short electrical traces, and suddenly your power-per-bit numbers look a lot better. Ayar's specific approach uses silicon photonics chiplets — they're doing optical I/O that can be integrated into a standard chiplet-style assembly, which means in principle you can mix their optical chiplets with compute chiplets in the same package using an interposer or advanced packaging techniques. Skyler: The funding signal here is important beyond just Ayar. It tells you that hyperscalers and system vendors are actively betting that co-packaged optics is not a research curiosity — it's a production roadmap item. Now, the timeline is still real. Getting co-packaged optics qualified into high-volume data center deployments requires solving yield, thermal, and reliability questions that pluggables have already solved over decades. But 650 million dollars buys you a lot of engineering time to work through those problems. This is one to watch closely — if Ayar lands a major hyperscaler design win at scale, it reshapes the optics supply chain. Skyler: Alright, story number two. The Next Platform has a piece on Marvell's A I business, and the headline is something I find genuinely interesting: optics is still driving Marvell's A I revenue more than custom silicon is. And I think that surprises people who've been watching the custom A I chip narrative dominate the conversation. Skyler: Here's the context. Marvell has two big A I plays. One is their custom A I silicon business — they're doing ASIC work for hyperscalers, helping them design custom training and inference chips, which is the same game Broadcom is playing. That's the business everyone talks about when they talk about the merchant silicon versus custom ASIC debate. But the other play is their optical DSP and electro-optics business, and apparently that's the one carrying more weight right now in terms of actual revenue contribution. Skyler: Think about what that means structurally. Every single A I cluster that gets built needs optics — lots of it. Whether you're connecting GPUs within a rack using direct attach copper or DAC, or you're going rack-to-rack, or you're going spine-to-leaf across a fat-tree topology, at some point light has to move data. And as cluster sizes scale from hundreds of GPUs to tens of thousands, the optical interconnect density goes up dramatically. Marvell sells the DSPs — the digital signal processors — that sit inside optical transceivers and coherent modules, handling the signal conditioning and forward error correction that makes those links work. Skyler: The reason this is interesting from a competitive standpoint is that optical DSP is a more stable business in some ways than custom silicon. Custom ASIC is lumpy — you land a big design win and you've got a three-to-five year program, but you've also got a lot of NRE up front and the revenue is concentrated. Optical DSP is more of a volume game: every transceiver needs a DSP, and when the datacenter is buying hundreds of thousands of ports, that adds up. Marvell has been quietly dominant in that space, competing with Coherent, InPhi before Marvell acquired them, and increasingly facing competition from in-house silicon at some of the hyperscalers. Skyler: The takeaway here for datacenter architects and investors: don't sleep on the optics supply chain as an A I infrastructure play. The custom chip narrative is real and it's growing, but the boring, essential, volume business of moving bits over glass is what's actually keeping the lights on at a lot of these semiconductor companies right now. Skyler: Story three — and this one is a bit of a left turn from the usual fabric coverage, but I think it belongs here. QuiX Quantum just announced the Carina system, which The Next Platform is covering as a significant step forward for photonic quantum computing. Now, quantum computing is not what this show is normally about, but let me tell you why I'm including it. Skyler: Photonic quantum computing is specifically interesting from an interconnect perspective because the qubit medium is light — photons. Unlike superconducting qubits, which need to be kept at millikelvin temperatures and are extremely sensitive to their environment, photonic qubits operate at room temperature and can potentially be transmitted over fiber. That last point is the one that matters for this audience. Quantum networking — using entangled photons to create secure communication channels or to link quantum processors — would, if it ever becomes practical, live on top of the same physical infrastructure as classical optical networking. Skyler: QuiX's Carina system is a photonic integrated circuit based quantum processor. The significance is in the integration density and the scale of the photonic chip — they're pushing more modes, more optical components, onto a single chip, which is the photonic equivalent of cramming more transistors onto a die. This is hard. Silicon photonics for classical networking is hard enough, but quantum photonics has additional constraints because you need to preserve quantum coherence and deal with photon loss in the waveguides. Skyler: Now, I'm not going to tell you that photonic quantum computing is about to land in your datacenter. It's not. The timelines for fault-tolerant quantum computation are still long and the engineering challenges are enormous. But here's why it's relevant: the photonic integration techniques being developed for quantum computing and the silicon photonics being developed for classical interconnect are increasingly cross-pollinating. The waveguide fabrication, the electro-optic modulator designs, the low-loss coupling techniques — these are shared problems. So when QuiX pushes photonic integration forward on the quantum side, there are real learnings that flow back to the classical side of the house, including the co-packaged optics work we just talked about. Skyler: Think of it as photonics having two frontiers right now, classical and quantum, and both of them are pushing the state of the art in integrated photonics fabrication. For the people in this audience who work in silicon photonics or optical interconnect, the quantum side is worth keeping an eye on even if you'll never touch a qubit. Skyler: Okay, story four — and this one comes from a slightly different angle. There's a piece making the rounds about copper demand surging in A I datacenters, framing it as a mining investment opportunity. And I want to talk about this not from the mining angle but from the infrastructure angle, because the copper story in A I clusters is real and it's underappreciated. Skyler: When people think about A I datacenter infrastructure, they think about GPUs and optics and high-speed fiber. And that's all real. But copper is everywhere. Direct attach copper cables — DAC cables — are the dominant interconnect for short distances inside a rack and between adjacent racks because they are cheap, they are low power, and they work great at distances up to about three to five meters. As you push to 400 gigabit and 800 gigabit speeds, DAC cables are getting harder to make work at those distances, which is why active copper cables — ACC — and active electrical cables are growing. Skyler: But beyond cables, there's copper in the PCB traces, in the package substrates, in the power delivery networks. And power delivery is a big one. The amount of copper required to deliver hundreds of kilowatts per rack — which is where A I compute is heading — is substantial. You need thick bus bars, you need heavy gauge power cabling, you need copper in the liquid cooling manifolds. It adds up. Skyler: Now, the mining investment angle I'll leave to others. But from a datacenter engineering perspective, the shift toward higher density A I compute is genuinely changing the material requirements. We're not just buying more of the same stuff — we're buying different configurations of copper infrastructure optimized for massive parallel power delivery and for very high bandwidth, very short reach electrical signaling. There's actually a real competition happening between copper and optics at that three to ten meter range, with active copper cables trying to hold off the optical invasion at short distances. Skyler: The copper versus optics boundary is a moving line, and it matters for how you design the physical layer of your A I fabric. Higher speeds push that boundary shorter — what copper could do at 100 gigabit per second, optics needs to take over sooner at 800 gigabit per second. So as the industry marches toward 1.6 terabit ports and beyond, expect optics to eat further into copper's traditional turf. But copper isn't going anywhere fast — it's cheaper, it's simpler, and the physics still favor it at the shortest distances. Skyler: Alright, last story, and this one is the one I've been wanting to talk about. The Ultra Ethernet Consortium — the UEC — has officially been announced with founding members that include leading cloud service providers, semiconductor companies, and system vendors, all coming together under the Linux Foundation. And this is a big deal for anyone who cares about how A I clusters are networked. Skyler: Let me back up and explain what problem the UEC is trying to solve, because the name is a little misleading. This is not about making Ethernet faster — we have the IEEE for that. What UEC is about is making Ethernet better for A I workloads specifically. And the reason that distinction matters is that traditional Ethernet was designed for a world where you have lots of different flows from lots of different sources, and occasional congestion is fine because TCP backs off and retransmits. A I training is completely different. You've got massive, synchronized, all-to-all communication patterns — collective operations like all-reduce — where every single endpoint needs to be talking to every other endpoint at high bandwidth, low latency, and with minimal jitter. If one flow gets congested and drops packets, you stall the entire training job. Skyler: The UEC is building a suite of specifications — transport enhancements, congestion control algorithms, endpoint behavior — that turn Ethernet into something that can actually compete with InfiniBand for A I workloads. And the fact that they're doing this under the Linux Foundation signals that the intent is an open, multi-vendor ecosystem, not a proprietary solution. This is a direct challenge to NVIDIA's InfiniBand dominance in A I cluster networking. InfiniBand has been the gold standard for high-performance computing and A I training because of its RDMA capabilities and its predictable, low-latency behavior. The Ethernet camp has been trying to close that gap for years. Skyler: Now, I want to be clear about where UEC specs actually stand: this is still a work in progress. The consortium has been announced, the working groups are forming, and the specifications are being drafted. We don't have a complete, ratified UEC standard that you can go implement today. What we have is a serious, well-funded, multi-stakeholder effort to produce one. The timeline from consortium formation to production-ready silicon is not months — it's a couple of years at minimum. So temper your expectations accordingly if you're a datacenter architect wondering whether to deploy this next quarter. Skyler: But the membership list matters. When you have hyperscalers, semiconductor companies, and system vendors all signing up together, that's not a paper standard — that's a supply chain alignment signal. These companies are telling their engineering teams: this is the direction, start building toward it. For the competitive landscape, it puts more pressure on NVIDIA to either open up InfiniBand more or make ROCEv2 and their own Ethernet solutions more competitive. It also creates a real opportunity for merchant silicon vendors who can build UEC-compliant switch and NIC silicon. This is a slow-moving wave but it's a real one. Skyler: And that's a wrap for today. We went deep on co-packaged optics funding, Marvell's optics-first A I business, photonic quantum computing's crossover with classical interconnect, copper's enduring role in A I fabric, and the Ultra Ethernet Consortium starting to take shape. A lot moving at once in this space — that's kind of always how it is. Line Rate, brought to you by Ada and Zen Semiconductor — Ada helps teams actually integrate A I into their workflows, and Zen Semiconductor is an umbrella company across the silicon industry. If you got something out of today's episode, please subscribe and share it with someone in your orbit who cares about this stuff — it genuinely helps the show grow. All the links and context for today's stories are in the show notes. I'm Skyler Monroe, thanks for listening, and I'll see you next time on Line Rate.