Molex to Acquire Teramount to Expand Co-Packaged Optics Capabilities - citybiz

Skyler: Hey, welcome back to Line Rate — the show tracking the chips, standards, and companies that move data at speed inside the datacenter, rack to rack, appliance to appliance, and beyond. I'm Skyler Monroe, and if you care about A I fabric and high-speed interconnect, you're in the right place. Big thanks to our sponsors Ada, who helps organizations actually integrate A I into their real workflows, and Zen Semiconductor, an umbrella company with a portfolio across the silicon industry. Couldn't do this without them. Alright — today we've got a lot happening in co-packaged optics land, and honestly it feels like the whole industry decided to make announcements in the same week. We've got Molex acquiring Teramount, Ayar Labs teaming up with Wiwynn, Sivers and GlobalFoundries moving on integrated photonics, Corning jumping into the CPO stack with GlobalFoundries, and Cisco making its case for Ethernet as the spine of A I supercomputing. Let's get into it. Skyler: First up — Molex is acquiring Teramount, and if you haven't been tracking Teramount, this is a good moment to pay attention. Teramount is an Israeli photonics company that's built a name for itself with what they call photonic wire bonding — a technique for coupling light between optical fibers and photonic integrated circuits with really tight tolerances and without the alignment nightmares that have historically made packaging photonics so painful and expensive. Skyler: So why does this matter for Molex? Molex is a connector and cable assembly company — they're deep in the physical layer, they're already inside every major datacenter with copper and fiber interconnects. But co-packaged optics is fundamentally a packaging problem, and the hardest part of that packaging problem is the fiber-to-chip interface. Getting light efficiently from a fiber into a silicon photonics die, at scale, with yield that makes economic sense — that's the unsolved or at least not-well-solved piece of the puzzle. Skyler: Teramount's approach essentially automates and tightens that coupling process. Their technology uses what they describe as a photonic funnel concept — a structure that relaxes the alignment tolerances so you don't need sub-micron precision at every connection point. For a company like Molex that's going to be assembling these things at volume, that's not a nice-to-have, that's a necessity. Skyler: The strategic read here is pretty clear. The CPO market is coming, the timeline has been sliding but it hasn't gone away, and the connectors-and-cable players know that if they don't own a piece of the optical coupling problem, they become a commodity supplier to whoever does. Molex is buying their way into a defensible position in the CPO stack. This is a smart vertical integration move. The acquisition price wasn't disclosed, but honestly the strategic value here is more interesting than the number anyway. Watch for Molex to start showing up in CPO reference designs and co-development agreements with switch ASIC vendors over the next couple of years. Skyler: Next — let's talk about Cisco's Ethernet pitch for A I supercomputing, because Cisco published a pretty detailed post laying out their argument and it's worth engaging with on the merits rather than just dismissing it as vendor marketing. Skyler: The core argument Cisco is making is that Ethernet — specifically high-speed Ethernet combined with the Ultra Ethernet Consortium's transport improvements — is the right fabric for large-scale A I training clusters, and that the InfiniBand versus Ethernet debate is tilting toward Ethernet as clusters scale into the tens of thousands of GPUs and beyond. Skyler: Now, regular listeners know I've covered U E C a bunch, and the honest picture is that the spec is still maturing. The transport layer work — things like out-of-order handling, congestion control mechanisms that actually work for A I all-to-all traffic patterns, the reliability layer — these are real improvements over vanilla Ethernet, but they're not fully baked in silicon yet. You've got early implementations, you've got the spec evolving, and the interoperability story is still being written. Skyler: That said, Cisco's underlying point about scale isn't wrong. When you're talking about a cluster that spans multiple racks, multiple pods, potentially multiple buildings, the operational tooling, the management plane, the people who know how to run it — Ethernet wins those arguments almost by default. The datacenter networking workforce knows Ethernet. The monitoring tools know Ethernet. The routing protocols know Ethernet. Skyler: Where I'd push back on the Cisco framing is on latency and tail latency specifically. InfiniBand's RDMA story is still cleaner for tightly coupled, latency-sensitive all-reduce workloads. The U E C transport work is trying to close that gap, and it's making real progress, but the gap isn't closed yet. The practical reality for most hyperscalers right now is a hybrid answer — InfiniBand or proprietary fabric for the tight compute fabric, Ethernet for the storage and east-west traffic. Cisco knows this, but they're making the long-game argument, and honestly the long game probably does go their way as U E C matures and the installed base of Ethernet at scale becomes the path of least resistance for new cluster builds. Skyler: Alright, story three — Sivers Semiconductors and GlobalFoundries announcing they're advancing A I datacenter optical solutions together. This one's interesting because it brings together two very different pieces of the photonics supply chain. Skyler: Sivers is a Swedish company that makes indium phosphide based lasers and optical components. GlobalFoundries is a leading contract semiconductor fab — not the leading-edge logic fab, but they have a really credible silicon photonics platform called GF Fotonix that they've been building out. The collaboration here is about integrating Sivers' laser technology with GF's silicon photonics process. Skyler: Here's the core technical problem they're solving: silicon photonics is great for routing light, modulating it, doing the waveguide work — but silicon is terrible at generating light. You cannot make a practical laser in silicon. So every silicon photonics platform needs an external laser source, and the way you integrate that laser with the silicon photonics chip is one of the critical engineering challenges in CPO and in optical transceivers generally. Skyler: Sivers does indium phosphide lasers, which are the workhorse of optical communications — high efficiency, good wavelength characteristics, they're what's inside most coherent transceivers. Integrating InP lasers onto or into a silicon photonics chip is not trivial — you're bonding dissimilar materials with different thermal expansion coefficients, different crystal structures. There are a few approaches: flip-chip bonding, wafer bonding, heterogeneous integration. This collaboration seems to be going after that integration problem at the process level, which is where you need to solve it if you want yield and volume. Skyler: For the CPO roadmap specifically, having a credible laser integration story is table stakes. You cannot do co-packaged optics without dealing with the laser source, and right now that's one of the messier parts of every CPO proposal out there. If Sivers and GlobalFoundries can make this clean and manufacturable at the GF fab, that's a meaningful contribution to the supply chain. Skyler: Story four — Ayar Labs and Wiwynn. This one got a little less attention than it deserved, so let's spend some time on it. Skyler: Ayar Labs has been one of the most technically ambitious companies in the CPO space. Their approach is different from most — instead of putting the optics at the edge of the package, they want to put optical I/O directly inside the package, very close to or even under the compute die. They call their technology TeraPHY — it's an optical I/O chiplet that can plug into a standard chiplet socket and provide optical connectivity that essentially replaces the electrical SerDes that would otherwise handle high-speed I/O. Skyler: The value proposition is significant. If you can replace SerDes with optical I/O inside the package, you cut the power consumed by I/O dramatically — Ayar claims something like five to ten times better energy efficiency per bit compared to electrical SerDes at the speeds we're talking about for A I training links. You also get much better bandwidth density and you don't have the reach limitations of copper at those data rates. Skyler: Now, Wiwynn is a Taiwanese ODM — one of the major server and rack-scale systems builders that supplies hyperscalers. They're not a household name outside the datacenter industry but they're significant. They build at scale, they work directly with the big cloud players. This partnership is Ayar Labs getting a serious systems integration partner who can actually take their optical I/O chiplet and figure out how it fits into real server and rack designs. Skyler: That's the gap Ayar needs to close. Having impressive chiplet technology is one thing. Getting it into a shipping product that a hyperscaler will actually deploy requires someone who knows how to build the thermal solution, the power delivery, the mechanical design, the supply chain. Wiwynn is that partner. This is a maturation signal for Ayar — they're moving from "cool chip on a demo board" toward something that looks like a real deployment path. Skyler: The competitive context here is that Ayar is going up against a lot of companies who are taking a more incremental approach to CPO — putting optics at the package edge rather than inside it. Ayar's inside-the-package story is higher risk, higher reward. Getting Wiwynn involved is a meaningful step toward validating the reward side of that bet. Skyler: And finally — GlobalFoundries and Corning. Yes, GlobalFoundries shows up twice this week, which tells you something about how active they're being in the CPO space right now. Skyler: This collaboration is specifically about fiber connectivity for co-packaged optics, and Corning is the right partner for that conversation. Corning basically invented modern optical fiber and they've been the dominant fiber manufacturer for decades. But this partnership isn't just about fiber — it's about how fiber connects to the CPO module at the system level. Skyler: One of the underappreciated challenges in CPO deployment is what happens at the point where fiber meets the package. In a traditional pluggable transceiver, you've got a standardized mechanical interface — a QSFP or OSFP connector — and the fiber plugs into that. It's clean, it's hot-swappable, the fiber routing is manageable. In a co-packaged optics design, the optics are integrated onto the switch package, which means fiber is connecting directly to something that's soldered onto the board. You lose hot-swappability. The fiber routing gets more complex. The connectors have to handle more precision because you're not going through a pluggable intermediary. Skyler: Corning's angle here is multi-fiber connectivity — specifically their work on multi-fiber push-on connectors and what they call edge-coupling solutions that can work at the tight geometries CPO demands. The collaboration with GlobalFoundries is about making sure the fiber interface side of GF's silicon photonics platform has a credible story. Because if you're a system designer evaluating CPO platforms, the laser integration question and the fiber connectivity question are both on your checklist, and GF is apparently working both of them simultaneously — Sivers for the laser side, Corning for the fiber side. Skyler: The big picture on all of this CPO activity is that the industry is clearly moving from "this is theoretically a good idea" to "let's actually build the supply chain." You've got packaging specialists, laser companies, fiber companies, ODMs, and fabs all making concrete moves in the same window. That's not coincidence, that's the industry responding to real demand signals from hyperscalers who have committed to CPO in their roadmaps. The technology has to be there when the demand lands, and right now everyone is racing to own a piece of the stack. Skyler: Alright, that's a wrap on today's Line Rate. We covered a lot of ground — from Molex buying into the CPO packaging stack with Teramount, to Cisco's Ethernet-for-A I argument, to Sivers and GlobalFoundries tackling the laser integration problem, to Ayar Labs and Wiwynn moving toward real deployment, and GlobalFoundries and Corning closing the loop on fiber connectivity. The through-line is that co-packaged optics is transitioning from a research narrative to an engineering execution problem, and the supply chain is mobilizing accordingly. If you want to go deeper on any of these, the show notes have links to all the sources. Line Rate, brought to you by Ada and Zen Semiconductor — Ada helps you integrate A I into your workflows, and Zen Semiconductor is building out a serious portfolio across the silicon industry. Subscribe wherever you get your podcasts, share with someone who needs to know this stuff, and I'll see you next time on Line Rate.