Lightmatter unveils Guide DR laser NIC to support CPO scale-up - Data Center Dynamics

Skyler: Hey, welcome to Line Rate — the show tracking everything happening in datacenter fabric and A I high-speed interconnect. I'm Skyler Monroe, and if it moves bits between accelerators, racks, or datacenters, we're covering it. Quick thanks to our sponsors: Ada, which helps organizations actually integrate A I into their workflows in meaningful ways, and Zen Semiconductor, an umbrella company with a portfolio of businesses across the silicon industry — including Ada on the A I deployment side. Alright, we've got a packed show today — a laser NIC from Lightmatter, Marvell going big on L P O at 1.6 terabits, VIAVI doing something interesting for Ultra Ethernet testing, a short seller taking aim at a data center firm, and a good explainer on rack-scale networking that's worth unpacking. Let's get into it. Skyler: Okay, story one. Lightmatter just unveiled something called the Guide DR — a laser NIC designed to support co-packaged optics scale-up. And if you've been following Lightmatter, this is a pretty significant product moment for them. Skyler: So first, let's set context. Lightmatter is probably best known for Passage, their photonic interconnect fabric — the idea being that you use photonics not just for I/O but as an actual compute interconnect layer. They've been making the case that silicon photonics and waveguide-based interconnect are the path forward for scaling A I clusters without hitting the power and bandwidth walls that copper is running into. Skyler: The Guide DR is a laser NIC, and the "DR" designation here is relevant — DR in the optics world typically refers to a reach and lane configuration, in this case pointing at short-reach, direct-attach scenarios. What Lightmatter is doing is building a NIC that has the laser integration baked in to support C P O — co-packaged optics — deployments at scale. Skyler: Now, why does this matter? Co-packaged optics is one of those things where everyone agrees it's coming, and almost everyone agrees the transition is going to be painful. The basic idea with C P O is that instead of pluggable transceivers sitting at the faceplate of a switch, you move the optical engine right next to — or even on top of — the switching ASIC. You get massive improvements in power efficiency and bandwidth density, but you lose the pluggability that the industry has basically built its operational model around for decades. Skyler: The laser source question is actually one of the thorniest parts of C P O. You need reliable, efficient, thermally stable lasers very close to the die, and historically that's been a hard problem. What Lightmatter is positioning Guide DR to do is provide that laser source as part of a NIC — so you're essentially building the optical drive capability into the host-side compute card, not just into the switch side. Skyler: For A I scale-up specifically — meaning the fabric that connects GPUs or A I accelerators within a job — this is pretty compelling. Scale-up traffic is brutal. It's all-to-all, it's latency sensitive, and it's running essentially continuously during training. Copper at 1.6T per port is getting very difficult. The signal integrity constraints are real, the reach is limited, and the power is high. Moving to optical for scale-up, with C P O as the endgame, is where the puck is going. Skyler: Lightmatter's angle here is interesting because they're not just selling a component — they're selling a vision of the whole system. Guide DR fits into their broader narrative about Passage being an interconnect fabric that's optically native from the ground up. Whether you buy that full vision or you just need a laser NIC for your C P O evaluation, this is a real product entering a space that has very few shipping solutions right now. Skyler: The competitive landscape here is intense. You've got Ayar Labs doing similar work on in-package optical I/O. You've got the big transceiver players — Coherent, II-VI, Lumentum — all trying to figure out their C P O play. And then you've got hyperscalers with their own internal programs. Lightmatter coming out with a specific product targeted at the laser NIC category is a way of staking territory before the C P O transition really accelerates. Skyler: My read: watch how this gets received by the hyperscaler and ODM community. If Guide DR starts showing up in evaluation programs at the big A I cloud providers, that's a signal that Lightmatter is genuinely in the race. The technology is real — the question is always timing and ecosystem readiness, and C P O has a history of timelines slipping. But the pressure from A I training workloads is real too, and that pressure is what actually makes things ship. Skyler: Story two is a bit different — we're stepping back from the pure technology and into the business and capital markets side, which honestly is just as important for understanding where this industry is going. Skyler: A short seller has published a report targeting Csquare, a data center firm, arguing that the company's shares are essentially worthless. The core of the argument, based on the reporting, centers on concerns about the portfolio — meaning the actual assets and contracts that underpin the business — and the capital structure, which is the financial architecture of how the company is funded and what it owes. Skyler: Now I want to be careful here because short seller reports are advocacy documents — they are written by people who have a financial stake in the stock going down. That doesn't mean they're wrong, but it means you read them the same way you'd read a sell-side initiation from a banker who's underwriting the deal. There's a perspective baked in. Skyler: That said, the framing around "portfolio and capital structure" is worth unpacking in the context of the broader data center market right now. We are in a moment where a lot of capital has flooded into data center development — driven by A I demand, driven by hyperscaler expansion, driven by the assumption that compute demand will grow faster than infrastructure can be built. That has created a wave of new entrants, special purpose vehicles, and data center companies that are leveraging up to build capacity in anticipation of demand that may or may not materialize at the assumed timeline or price. Skyler: The short thesis on data center companies, in general, often comes back to a few things: are the contracts actually signed and creditworthy? What's the cost of capital relative to the yield on the assets? And how much of the growth story is priced in at a level that assumes perfect execution? Skyler: For listeners in the infrastructure and hardware space, this is relevant because capital markets are part of the ecosystem. If data center developers — especially the ones that aren't the Tier 1 hyperscalers — get into financial stress, that affects procurement cycles, it affects which hardware gets ordered and when, and it affects the rollout of some of the next-gen fabric technology we talk about on this show. A strained data center developer is not deploying 1.6T optics or next-gen switches on an aggressive timeline. Skyler: I'm not going to weigh in on whether Csquare specifically is overvalued or not — I don't have enough visibility into their books and I'm not in that business. But the existence of this kind of short attack in the data center space is a signal that the market is starting to apply more scrutiny to the A I infrastructure buildout thesis, which has mostly been treated as a free pass for the last two years. Skyler: The broader point: not every data center company riding the A I wave has the same underlying asset quality or balance sheet resilience. As capital costs stay elevated and the timing of A I demand becomes clearer — or murkier — expect more differentiation between the strong operators and the ones that were basically structured finance vehicles dressed up as infrastructure companies. Skyler: Story three. The Register published an introduction to rack-scale networking, and I actually want to spend some time on this because explainers matter — they shape how people think about the architecture, and there are some important conceptual points here that are worth getting right. Skyler: So what is rack-scale networking? At the highest level, it's the idea that the fundamental unit of infrastructure is no longer a single server or a single switch — it's the rack. You design the networking, power, and compute as an integrated system at the rack level, and you optimize across all of those dimensions together rather than independently. Skyler: This sounds almost obvious when you say it out loud, but it represents a real departure from how data centers have traditionally been built. The classic model is: buy servers, buy switches, cable them together with pluggable transceivers per some topology, and manage the compute and networking layers somewhat independently. Rack-scale challenges that by asking: what if we made the networking first-class at the rack level from day one? Skyler: For A I specifically, this is really important. A I training clusters are not general purpose compute. They're doing a specific thing — running collective operations, all-reduce, all-to-all, across many accelerators simultaneously — and the networking requirements are extremely specific. High bandwidth, low latency, very predictable, and often a fixed topology because the job doesn't move once it starts. That profile is very different from, say, a web serving workload where you want flexible east-west switching. Skyler: Rack-scale networking for A I tends to manifest in a few different ways. You've got direct attach copper or optical cables within the rack connecting GPUs or accelerators to a top-of-rack switch or to each other. You've got the emergence of NVLink switches and similar proprietary fabrics that do scale-up within a rack or a small number of racks. And then you've got the scale-out layer that connects racks to each other — which is where Ethernet and InfiniBand are competing. Skyler: One of the things the explainer touches on is the distinction between scale-up and scale-out, which I think is the most important architectural split in A I networking right now. Scale-up is the tight, low-latency fabric that connects accelerators within a job — think NVLink, or the C X L and U A Link space as it matures. Scale-out is the fabric that connects jobs to each other, connects to storage, handles the broader cluster routing — that's where Ethernet and InfiniBand play. Skyler: The interesting design question at the rack level is where you draw that line. Some designs keep scale-up entirely within a single rack — 72 GPUs on NVLink, full mesh within the rack, then Ethernet out. Others push scale-up across multiple racks using high-speed optical links, which is where you start needing the kind of optics we talked about with Lightmatter. Skyler: Rack-scale also implies something about power architecture — high-density racks for A I are running 30, 40, 60, even 100 kilowatts per rack now in some deployments, and the networking infrastructure has to be designed with that thermal reality in mind. Pluggable transceivers burn power. Active copper burns power. The push toward C P O and on-board optics is partly about fitting more bandwidth into a rack power budget. Skyler: If you're a networking engineer or architect and you haven't thought through your rack-scale topology choices explicitly, now is the time. The decisions you make at the rack level — topology, cabling, switch placement, optics choice — have enormous downstream consequences for your A I cluster performance, your power costs, and your ability to scale. Skyler: Story four. Marvell has announced a 1.6 terabits per second L P O chipset targeting optical short-reach, scale-up compute fabric interconnects. This is a significant product announcement and I want to dig into what L P O actually means and why Marvell is betting on it. Skyler: L P O stands for linear pluggable optic. It's a design philosophy for optical transceivers that strips out the DSP — the digital signal processor — that you'd normally find in a traditional pluggable. In a conventional C F P or Q S F P transceiver, you've got a DSP doing retiming, equalization, and FEC — forward error correction — on both the electrical and optical sides. That DSP is doing a lot of work but it also consumes a lot of power and adds latency. Skyler: L P O basically says: what if we trusted the signal enough to skip the DSP entirely, or at least make it very thin? You connect the host-side SerDes directly to the optical engine with minimal processing in between. The result is lower power, lower latency, and lower cost — but it puts more demands on the electrical signal integrity on the host side, and it means you need very good SerDes performance in the ASIC. Skyler: At 1.6T, you're talking about — depending on the lane configuration — something like 16 lanes at 100 gigabits per second per lane, or 8 lanes at 200G. Either way, this is at the absolute frontier of what electrical SerDes can do. The signal integrity at those speeds is genuinely hard, and the reason L P O becomes attractive at this data rate is that a DSP-based approach at 1.6T would be ferociously power hungry. Skyler: Marvell is specifically calling this out for scale-up compute fabric. So we're back to that same architectural point from the rack-scale discussion — this is about the tight interconnect between accelerators, not the general-purpose Ethernet switching fabric. At 1.6T, you're building the links that are going to connect next-generation GPU clusters in ways that push beyond what NVLink currently does at the inter-rack level, or that give Ethernet-based scale-up a fighting chance against proprietary solutions. Skyler: The L P O versus C P O framing is interesting here. L P O keeps pluggability — you still have a module you can swap. C P O removes pluggability entirely in exchange for even better power and density. Some in the industry think L P O is a bridge — a way to get some of the power benefits while the ecosystem catches up to full C P O. Others think L P O will have a longer life than people expect because operations teams really value being able to swap a transceiver without touching the switch ASIC. Skyler: Marvell is well positioned to play in this space because they have strong SerDes IP and a history of building the guts of networking ASICs that end up in a huge range of customer designs. This isn't a merchant silicon company trying to sell you a switch chip — Marvell is often the company behind the ASIC or the PHY that someone else's branded product is built on. Their 1.6T L P O chipset is the kind of thing that could end up in products from multiple switch vendors or hyperscaler custom designs without ever showing up prominently in a press release. Skyler: The competitive tension here is with companies like Broadcom, which has its own approach to high-speed SerDes and optics integration, and with the pure-play optics guys who are also developing L P O-compatible solutions. Marvell's play is to own the electrical side of the interface — the part that talks to the optics — and make that so good that the L P O transceiver market scales around their chipset. Skyler: For A I fabric architects, 1.6T per port in an L P O format changes the calculus on scale-up topology significantly. You can build fatter pipes with fewer hops, which is exactly what collective operations want. Watch how this gets adopted by the ODM and system integrator community over the next 12 months. Skyler: Last story. VIAVI Solutions has launched what they're calling the industry's first validation solution for Ultra Ethernet Transport — U E T — and they're positioning it as accelerating A I data center deployments. This is a testing and validation story, and those matter a lot more than they get credit for. Skyler: Quick context on Ultra Ethernet. The Ultra Ethernet Consortium — U E C — was formed in 2023 with a pretty serious roster: AMD, Arista, Broadcom, Cisco, Eviden, HPE, Intel, Meta, Microsoft, and others. The goal is to define an Ethernet-based protocol stack optimized for A I and H P C workloads — one that addresses the specific pain points that traditional Ethernet has in the collective operations world. Things like congestion control, out-of-order delivery, and the fact that standard Ethernet was designed for TCP/IP traffic patterns that look nothing like an all-reduce across 10,000 GPUs. Skyler: U E T — Ultra Ethernet Transport — is the transport layer specification within that effort. It's the part that handles how packets move reliably and efficiently over the Ultra Ethernet fabric. And I want to be clear: U E C specs are still evolving. This is not a fully baked, ratified, ship-your-product-against-it standard yet. It's moving fast, there's industry momentum, but you should be aware that details can and do change as the working groups iterate. Skyler: So into that environment, VIAVI is launching a test and validation solution specifically for U E T. And this is actually a really important kind of product in the standards development cycle. Here's why: when you're building networking equipment against a spec that's still being refined, you need tooling that can verify your implementation is doing what the spec says — and more importantly, tooling that can help you understand when your implementation diverges from what other vendors are doing. Skyler: Interoperability is the thing that makes standards real. A spec that exists on paper but where no two implementations actually work together is not a useful standard. Test equipment that exercises the spec — including the edge cases, the error conditions, the congestion scenarios — is what turns a spec into something you can actually build products around. Skyler: VIAVI calling this the "first" validation solution for U