Light into data: How silicon photonics is powering the AI data center revolution - STMicroelectronics

Skyler: Welcome to Line Rate — the show about datacenter fabric and AI high-speed interconnect. I'm Skyler Monroe, and if it moves bits between GPUs, across a rack, or between buildings, we're probably talking about it here. 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. Really appreciate their support. Alright — today we've got a lot of light. Silicon photonics, co-packaged optics, and Broadcom throwing a pretty direct punch at Nvidia. Let's get into it. Skyler: First up — STMicroelectronics published a piece this week on how silicon photonics is powering the A I datacenter revolution, and look, I know that sounds like marketing copy, but there's actually some real substance worth unpacking here. So let's talk about where we are with silicon photonics and why it keeps showing up in every serious conversation about next-generation fabric. Skyler: The core problem that silicon photonics is trying to solve is electrical signaling distance and power. When you're running copper traces — whether that's on a board or in a cable — you're fighting signal degradation and you're burning a lot of power in SerDes just to recover the signal at the other end. And as lane rates push toward 112 gig and now 224 gig per lane, that fight gets harder and more expensive in watts per bit. Skyler: Silicon photonics flips that around. You convert the electrical signal to light, move it over fiber — which has dramatically better loss characteristics — and convert it back. The challenge has always been making that conversion cheap, low-power, and manufacturable at scale. And this is where ST's piece is actually interesting, because they're talking about their own BiCMOS process integration, which lets them put the photonic components and the driving electronics on the same wafer, or at least in very close proximity. That's a meaningful manufacturing advantage. Skyler: Now why does this matter for fabric? Think about the architecture of a large A I training cluster — you've got hundreds or thousands of accelerators that need to talk to each other with very low latency and very high bandwidth. The electrical interconnect budget, meaning how far you can run copper before signal integrity falls apart, is measured in centimeters to maybe a meter or two at high lane rates. Once you go photonic, that budget opens up to kilometers if you need it. But even at rack scale, the power savings per bit are compelling. Skyler: ST is positioning themselves as a supplier to the ecosystem here — they're not building the switch chips or the A I accelerators, they're supplying the photonic components and process technology that goes into the optical engines. And that's a smart place to be right now because the demand signal from hyperscalers and A I companies is extremely strong and relatively supplier-agnostic at the component level. Skyler: The interesting question is where silicon photonics ultimately lands in the packaging hierarchy. You've got pluggable optics, which are mature and hot-swappable. You've got co-packaged optics, which we'll talk more about in a few minutes. And then you've got the longer-term vision of on-chip photonics, where the laser is literally integrated with the compute die. Each step gets you more bandwidth density and better power efficiency, but also more manufacturing complexity and less flexibility. ST is positioning their silicon photonics platform to serve across that spectrum, which is the right call at this stage of the market. Skyler: The bottom line: silicon photonics is not a future technology anymore. It is the technology that A I fabric is being designed around. And having more suppliers with real manufacturing capability — not just lab demos — is genuinely important for the supply chain. Skyler: Okay, next story, and this one has some teeth. Broadcom has announced Thor Ultra, an 800 gigabit Ethernet chip aimed squarely at A I cluster networking. And the way this is being positioned — open, standards-based Ethernet — is a very deliberate contrast to what Nvidia is selling with InfiniBand and now NVLink-based fabrics. Skyler: Let me give you the technical picture first. 800G Ethernet at the switch or NIC level means you're aggregating multiple high-speed lanes to get to that 800 gigabit per second figure. At 112 gig per lane, that's roughly 8 lanes per port. At 224 gig per lane — which is where the industry is heading with 224G SerDes — you can get to 800G with 4 lanes, which is a big deal for die area and power. Broadcom has been investing heavily in their SerDes technology, and Thor Ultra is presumably leveraging their latest work there. Skyler: Now, the "open" framing is doing a lot of work in this announcement. Broadcom is very deliberately invoking the Ultra Ethernet Consortium — the U E C — which is the industry consortium that's been working to extend standard Ethernet to be more competitive for A I training workloads. The traditional knock on Ethernet for A I training is that it wasn't designed for the all-to-all communication patterns you see in collective operations like AllReduce. InfiniBand's RDMA and congestion control were purpose-built for that. U E C is trying to close that gap with transport-layer improvements that stay within the Ethernet standards world. Skyler: Thor Ultra's relevance here is that it's a real chip, from a real company, that hyperscalers and OEMs can actually buy and build systems around. Consortium specs are great, but they need silicon to mean anything. And Broadcom shipping hardware that implements open standards is a meaningful forcing function on the ecosystem. Skyler: The competitive angle is obvious. Nvidia's NVLink 5 fabric, which connects H100s and B200s within a server node and between nodes via NVSwitch, is proprietary. It's extremely high bandwidth and low latency, and if you're building a cluster with Nvidia GPUs, you're probably using it. But a lot of hyperscalers and the people building A I infrastructure for alternative accelerators — AMD, Intel, various startups — need a fabric story that isn't tied to one vendor. That's the opening Broadcom is pushing into. Skyler: There's also a UALink angle worth mentioning. UALink — U A Link — is the consortium working on the scale-up fabric standard, meaning the high-bandwidth, low-latency connections between accelerators within a server node. Broadcom is active in that ecosystem too. Thor Ultra looks more like a scale-out play — rack-to-rack and cluster-wide — but the scale-up and scale-out worlds are increasingly connected in how people design systems. You need coherent answers for both. Skyler: The thing I keep watching with U E C is transport layer maturity. The spec is still evolving. Congestion control for A I collectives over Ethernet is hard — you need to handle incast, you need low-latency retransmit, and you need to do this at scale without blowing your tail latency. Broadcom shipping silicon is a step, but the software and firmware stack around it matters just as much. That part is still maturing. Skyler: If you're a datacenter architect evaluating fabric right now, Thor Ultra is legitimately worth putting on your test list alongside Nvidia's offerings. Broadcom's execution track record is strong, and they have deep relationships with every major cloud provider. This is a serious product. Skyler: Alright, pivoting to the optics investment side. Coherent — ticker COHR — had a Seeking Alpha piece this week making the bull case for the stock based on A I datacenter optics demand. I want to talk about Coherent specifically because they're one of the most interesting companies in the optics supply chain and their positioning tells you something about where the market is going. Skyler: Coherent is not a household name outside of the industry, but they are one of the largest vertically integrated optical component and module makers in the world. They do everything from indium phosphide wafers — which is the substrate for a lot of high-performance lasers — to finished transceiver modules to coherent optical systems for long-haul networking. That vertical integration is increasingly a strategic advantage. Skyler: The demand driver is pretty straightforward. Every 400G, 800G, and soon 1.6T port in an A I datacenter needs optical transceivers. As cluster sizes grow and lane rates increase, the transceiver count goes up and the technical requirements get harder. You need lower power per bit, higher density, and increasingly you're looking at co-packaged configurations where the optics are physically close to the switch ASIC rather than in a pluggable cage on the faceplate. Skyler: Coherent's revenue mix has been shifting toward datacom — meaning datacenter optics — and away from telecom. That's a deliberate strategy, and the margins in datacom, especially for higher-speed products, are attractive. The Seeking Alpha thesis is basically: A I capex is going up, optical content per server and per switch is going up, and Coherent has the technology and manufacturing scale to capture a large share of that spend. Skyler: What's technically interesting about Coherent in particular is their work on 1.6 terabit per second transceivers and their involvement in co-packaged optics development. Getting to 1.6T in a pluggable form factor is genuinely hard — you're pushing the boundaries of what you can do with pluggable modules in terms of power dissipation and signal integrity. Coherent has been one of the companies working on the optical engine architectures that make co-packaged optics viable. Skyler: From an industry structure standpoint, the optics supply chain has historically been pretty fragmented and competitive — lots of suppliers, commoditized at the lower speeds. But as you move up to 800G and 1.6T, the technical barriers go up significantly, and the number of suppliers who can actually deliver at volume shrinks. That dynamic tends to be good for the incumbents who have invested in the right technology. Coherent, II-VI before it became Coherent, has been making those investments for a while. Skyler: I'm not here to give investment advice — this is a technology show — but from a pure technology lens, Coherent is one of the companies whose success is structurally tied to A I fabric scaling, and that's a pretty good place to be right now. Skyler: Next up — AOI, that's Applied Optoelectronics Inc, put out a piece in their newsroom laying out their view of the future of A I datacenter networking across three categories: CPO, XPO, and what they're calling optical innovation more broadly. This is a useful framework, so let me walk through it because I think it clarifies some terminology that gets muddled. Skyler: CPO is co-packaged optics. The idea is that instead of a pluggable transceiver module that connects to the switch chip via a PCB trace and a connector, you package the optical engine directly with the switch ASIC — either in the same package or extremely close to it. This eliminates the electrical trace between the SerDes output and the optical modulator, which at 224 gig per lane is a meaningful signal integrity win. It also opens up the package bandwidth budget, because you're not constrained by the faceplate connector density anymore. Skyler: XPO — and this is the term AOI is using, though you'll see variations — refers to external photonics or near-package optics. It's kind of an intermediate point on the integration spectrum. The optics aren't literally in the same package as the switch ASIC, but they're very close — think a separate photonic die on the same substrate, or a small optical module attached directly to the ASIC package. You get some of the electrical benefits of co-packaging without the full manufacturing complexity of integrating everything into one package. Skyler: AOI's framing here is that these aren't sequential generations where one replaces the other — they're different solutions for different deployment scenarios. Pluggable is still king where you need flexibility and serviceability. CPO makes sense for high-density switch builds where power and bandwidth density are the primary constraints. XPO sits in the middle. That's actually a pretty mature and nuanced view of the market. Skyler: From an AOI-specific perspective, they're a smaller player compared to Coherent or Lumentum, but they've been competitive in the 100G and 400G datacom market and they're making a push into the higher-speed segments. Their newsroom piece is partly a positioning play — here's how we think about the technology landscape, and here's where we fit. But the framework they're laying out is legitimate and worth understanding. Skyler: The broader point about optical innovation is that the transition from pluggable to co-packaged is not primarily a standards battle — it's a manufacturing and supply chain challenge. The optical engines themselves need to be manufacturable at the yields and cost points that make them viable compared to pluggable alternatives. That's a hard problem, and it's one reason the CPO transition is taking longer than some early forecasts predicted. But the direction is clear. Skyler: For datacenter architects, the practical implication is that you're probably going to be living in a mixed world for longer than you'd like — some pluggable, some near-package, maybe some co-packaged in the highest-density switches — and your fabric design needs to accommodate that heterogeneity. AOI's point that these coexist rather than compete directly is actually operationally useful framing. Skyler: Alright, last story, and this one's a little different. MicroVision — yes, that MicroVision, the lidar company — announced something called the ELSFP module targeting co-packaged optics applications for A I datacenters. And I have to be honest with you, this one caught me off guard. Skyler: So MicroVision has historically been a lidar and sensing company. Automotive sensing, that kind of thing. The ELSFP module is apparently a fiber interface product designed to work in co-packaged optics architectures. The "ELSFP" name suggests it's some kind of enhanced small form-factor pluggable variant, though the details in the announcement are a bit thin on the technical specifics. Skyler: Here's the way I'd contextualize this. When a market gets hot — and co-packaged optics and A I datacenter optics are very hot right now — you start to see companies from adjacent markets trying to pivot or extend their product lines to capture some of that demand. MicroVision has optical and photonic expertise from their sensing work, and some of that knowledge base transfers. Laser technology, detector technology, optical alignment — these are shared disciplines. Skyler: Whether the ELSFP module is a serious CPO product or more of a market exploration play is hard to say from the announcement alone. The co-packaged optics market has some very serious, well-capitalized incumbents — your Coherents, your Lumantums, your Marvell and Broadcom for the ASIC side. Breaking into that requires not just a working product but qualification at hyperscalers, which is a long and expensive process. Skyler: What I'd watch for is whether MicroVision gets any design wins or letters of intent from actual datacenter customers. An announcement of a product is one thing. Getting onto the approved vendor list at a major cloud provider is a completely different level of validation. Skyler: There's a broader pattern here worth noting. The A I datacenter buildout is creating demand signals strong enough to pull in companies from adjacent markets — automotive sensing, industrial imaging, scientific instruments. Some of those pivots will be successful because the underlying technology really does transfer. Others will be opportunistic and won't find traction. It's hard to tell which is which at the announcement stage. Skyler: For the established players in the optics supply chain, I think they're watching these entrants but not losing sleep yet. The real competitive threats are from companies that already have hyperscaler relationships and are moving up the stack — or from accelerator vendors who might decide to vertically integrate their optical supply chain. Those are the scenarios that would actually reshape the market. Skyler: Okay, that's a wrap on today's stories. Big theme today was obviously light — silicon photonics, co-packaged optics, XPO, 800G Ethernet, and the companies trying to capture share of an optical supply chain that's scaling hard and fast. The throughline across all of it is that A I fabric is driving an enormous amount of investment and innovation in how we move bits — whether that's with photons, with open Ethernet silicon, or with new companies bringing in optical expertise from unexpected places. Skyler: Thanks for spending time with me on Line Rate, brought to you by Ada and Zen Semiconductor. Ada helps companies integrate A I into their workflows, and Zen is an umbrella organization with a portfolio of businesses across the silicon industry. Genuinely appreciate their support in keeping this show going. If you found today's episode useful, subscribe wherever you get your podcasts, share it with a colleague who cares about this stuff — and there are show notes with links to all the stories we covered today. See you next time.