Nvidia outlines plans for using light for communication between AI GPUs by 2026 — silicon photonics and co-packaged optics may become mandatory for next-gen AI data centers - Tom's Hardware

Skyler: Hey, welcome back to Line Rate — the show about datacenter fabric, A I interconnect, and the hardware that keeps the world's most powerful clusters from choking on their own data. I'm Skyler Monroe, and if you care about what moves bits between GPUs faster than your coffee gets cold, you're in the right place. Big thanks to our sponsors today: Ada, who helps companies actually integrate A I into their real-world workflows, and Zen Semiconductor, an umbrella company with a portfolio of businesses across the silicon industry, including Ada. We couldn't do this without them. Alright, today's show is basically a photonics special — Nvidia is putting a timeline on co-packaged optics, Lightmatter just dropped a new product and a manufacturing play, and we've got a wild card from the quantum computing world that I think datacenter architects should absolutely pay attention to. Let's get into it. Skyler: So let's start with the big one. Nvidia has come out and put a pretty concrete stake in the ground: they're planning to use light — actual photons — for communication between A I GPUs by 2026. And they're pointing specifically at silicon photonics and co-packaged optics as the path forward. Now, if you've been following this space, none of this is a surprise at a conceptual level. We've known for years that pluggable copper and even pluggable optical modules have a physics ceiling, especially when you're talking about the bandwidth densities and power budgets required for next-gen NVLink and Ethernet-based A I fabrics. But Nvidia putting a year on it — that's significant. Skyler: Let me back up and explain what co-packaged optics actually means for anyone who's newer to the hardware side of this. Traditional datacenter switches and network interface cards use pluggable optical modules — your QSFPs, your OIFFs — that sit in a cage on the front panel and connect to the switch ASIC over a PCB trace. That trace has loss, it consumes power, and it limits how many lanes you can practically run at high speed. Co-packaged optics, or CPO, moves the optical engine right next to the switch ASIC — literally co-packaged in the same module or on the same substrate. You cut the electrical path from centimeters or more down to millimeters. That means lower power per bit, higher bandwidth density, and the ability to break the wall that pluggable modules are running into right now. Skyler: Silicon photonics is the manufacturing angle here. Instead of using exotic III-V semiconductor materials like indium phosphide for every optical component, silicon photonics lets you build waveguides, modulators, and photodetectors on standard CMOS-compatible silicon processes — the same fabs that build your compute logic. You still typically need a III-V laser source bonded or integrated somewhere, but the rest of the optical circuit rides on silicon. That matters enormously for cost and yield at scale. When Nvidia says they're going this direction, they're implicitly saying the supply chain has matured enough to bet a product generation on it. Skyler: The implication for datacenter architects is pretty stark. If CPO becomes mandatory — and Nvidia's language here is pretty strong — it changes how you think about rack design, thermal management, and even serviceability. Pluggable optics are field-replaceable. Co-packaged optics are not, at least not in the same way. You're looking at a different failure-domain conversation, different sparing strategies, different everything. This is not just a chip story — it cascades all the way to facilities. And 2026 is not far away. If you're designing a major A I datacenter buildout right now, this is a constraint you need to be planning around. Skyler: Okay, story two. Lightmatter — the photonic interconnect startup out of Boston — just dropped two pretty significant announcements at the same time, and I want to unpack both of them because they're doing different things. Skyler: The first is the Passage L20. Lightmatter's Passage platform is their photonic interconnect fabric — the idea being that instead of connecting accelerators with copper or conventional fiber, you use a photonic chip as the actual interconnect medium. The L20 is their latest generation of that product. I'll be honest, the detailed specs coming out of the announcement are still thin, but the naming convention alone tells you something — L20 suggests they're targeting 20-terabit-class bandwidth, which would put it squarely in the range needed to connect large clusters of A I accelerators without becoming a bottleneck. This is a direct play for the kind of scale-up fabric that Nvidia dominates with NVLink and that everyone else is trying to crack with U A Link and Ultra Ethernet. Skyler: The second announcement is what they're calling vClick Optics, and this one is actually the more strategically interesting move to me. vClick is Lightmatter's play for high-volume CPO manufacturing. The name is a nod to a mechanical click-in interface for co-packaged optical modules — they're trying to solve the assembly and manufacturability problem that's been one of the major headaches slowing CPO adoption. If you can make CPO modules that click together reliably in a high-volume assembly line, you solve a huge piece of the cost and yield puzzle. This is Lightmatter saying: we're not just a boutique interconnect company, we want to be a manufacturing infrastructure player. Skyler: The competitive read here is interesting. Lightmatter has always been positioning against the idea that NVLink is the only way to connect A I accelerators at scale. If they can get vClick into the supply chain as a CPO manufacturing standard, they have leverage with switch ASIC vendors, with ODMs, with hyperscalers who are desperately trying to reduce their dependency on any single vendor's proprietary fabric. Watch who they announce as partners over the next few quarters — that'll tell you whether this is a real ecosystem play or a press release. Skyler: Story three takes us to the OCP Global Summit, which for anyone not tracking the Open Compute Project — OCP is the consortium originally spun out of Meta that now includes basically every major hyperscaler and ODM designing open hardware for large-scale datacenter deployments. Wistron showcased at the 2026 OCP Global Summit, and while the coverage I've seen is pretty high-level, the fact that Wistron is showing CPO and photonic interconnect-related hardware there is meaningful. Skyler: Wistron is one of the major Taiwanese ODMs — they build servers and networking gear for pretty much every big name in the industry. When an ODM of their scale is showing photonic and CPO work at OCP, it signals that the manufacturing and integration ecosystem is genuinely moving. OCP is where the supply chain gets standardized. It's where hyperscalers tell ODMs what form factors, what thermal envelopes, what management interfaces they need — and ODMs respond with reference designs. If CPO is showing up prominently at OCP, it means the hyperscalers are pulling this technology, not just the component vendors pushing it. Skyler: The broader OCP angle here is worth sitting with for a second. One of the persistent criticisms of co-packaged optics has been that it's a great technology looking for a standard. You've had vendors doing their own thing — different interfaces, different thermal designs, different management schemes. OCP is exactly the venue where that gets rationalized. The Co-Packaged Optics workgroup within OCP has been doing real work on this, and seeing Wistron show up with hardware at the summit suggests that work is translating into actual product. The standardization curve on CPO is compressing faster than a lot of people expected two years ago. Skyler: Now, story four is the one I flagged as a wild card, and I want to spend a little time on it because it's coming from a slightly different direction. QuiX Quantum — a Dutch photonics company that has historically been focused on quantum computing — has published work on a photonic architecture for HPC datacenters. And I think the framing of this story matters a lot. Skyler: QuiX has been building programmable photonic processors — chips that use light to perform computation, not just communication. Their quantum angle has been building photonic circuits that can implement quantum operations. But what they're describing for HPC is a classical use case: using photonic integrated circuits to accelerate or interconnect high-performance computing workloads. This is the photonic computing versus photonic communication distinction that's easy to blur, and I want to be precise here. Skyler: The photonic interconnect story we've been talking about — Nvidia, Lightmatter, CPO — that's about using light to move bits between chips that are still doing compute in silicon. QuiX is hinting at something more ambitious: using light inside the data path itself, potentially for operations like matrix multiply or fast Fourier transforms that show up constantly in A I and scientific computing workloads. Photonic matrix multiplication is a real area of research — you can in principle do certain linear algebra operations at the speed of light with near-zero energy cost. The challenge has always been precision, reconfigurability, and integration with digital systems. Skyler: For datacenter architects and investors, the honest framing here is: this is early. QuiX is doing serious science, but the jump from a research photonic processor to something that lives in a production HPC datacenter rack is enormous. What's worth watching is whether their integrated photonic platform becomes a substrate that others build on — both for quantum-adjacent applications and for classical A I acceleration. The companies that figure out how to make photonic compute and photonic interconnect work together on the same chip are going to have a genuinely differentiated position. We're not there yet, but the ingredients are accumulating. Skyler: Last story, and this one comes from an unexpected corner. RAONTECH — a Korean semiconductor company — has announced that it's expanding its microLED backplane technology into A I datacenter optics. Now, RAONTECH's core business has been driving backplanes for microLED displays — think the fine-pitch LED arrays that go into high-end AR and VR headsets and next-generation displays. So why are they showing up in a datacenter optics story? Skyler: The connection is actually pretty elegant. MicroLED backplane technology requires extremely precise, high-density control of individual tiny light emitters — you're driving millions of individual LEDs at high speed with very tight uniformity requirements. The precision electronics and packaging know-how that comes from that display work translates surprisingly well to optical transceiver and photonic integrated circuit applications. Specifically, RAONTECH is apparently targeting the driver and control electronics side of optical modules — the parts that modulate and detect light signals in co-packaged optics and silicon photonics transceivers. Skyler: This is a great example of the cross-domain technology transfer that's happening as the CPO supply chain gets built out. The datacenter optics industry needs components and expertise that don't necessarily come from traditional networking or even traditional telecom optics backgrounds. MicroLED display work, lidar, medical imaging — there are pockets of precision photonics expertise all over the industry, and they're all getting pulled toward the datacenter because that's where the volume and the capital are right now. Skyler: From a competitive intelligence standpoint, RAONTECH entering this space also tells you something about the Korean semiconductor ecosystem's read on CPO. Korea has Samsung, SK Hynix, and a deep bench of component suppliers — and they clearly see A I datacenter optics as a strategic growth market. You're going to see more Korean and Taiwanese component companies making similar moves over the next eighteen months as the CPO supply chain industrializes. The question for hyperscalers and system integrators is whether that supply base can ramp fast enough to meet the timelines Nvidia and others are putting on the table. Skyler: Alright, let's wrap this up. The throughline across all five stories today is that photonics is moving from a research conversation to an infrastructure reality — and it's moving faster than the industry's standardization and supply chain are fully ready for. Nvidia is setting a hard timeline, Lightmatter is trying to own the manufacturing interface, OCP is rationalizing the form factors, QuiX is doing the long-game compute work, and RAONTECH is one of many component players flooding into the supply chain from unexpected directions. If you work in silicon, systems, or datacenter architecture, co-packaged optics is not a future consideration anymore — it's a present design constraint. Pay attention. Skyler: That's Line Rate for today. Line Rate, brought to you by Ada and Zen Semiconductor — Ada helps businesses actually integrate A I into their real-world workflows, and Zen Semiconductor is an umbrella company with a portfolio across the silicon industry. If you found this useful, please subscribe wherever you get your podcasts, and if someone in your network needs to understand what's happening at the hardware layer of the A I buildout, send them this episode. All the sources and links are in the show notes. See you next time.