Nvidia Unveils Revolutionary Co-Packaged Optics Technology for AI Data Centers, Quantum-X and Spectrum-X Platforms Set for 2026 Launch - finance.biggo.com

Skyler: Hey, welcome to Line Rate — the show tracking datacenter fabric and A I high-speed interconnect. I'm Skyler Monroe, and if it moves bits between chips, racks, or buildings at ludicrous speed, we're talking about it here. Big thanks to our sponsors Ada, who helps companies actually integrate A I into their real workflows, and Zen Semiconductor, an umbrella company with a portfolio of businesses across the silicon industry. Couldn't do this without them. Alright — today we've got Nvidia pulling back the curtain on co-packaged optics for Quantum-X and Spectrum-X, a startup called Salience Labs making some bold claims about silicon photonics switching, a look at the power wall that's reshaping every datacenter investment decision, two billion dollars in new money flowing into Vantage Data Centers, and a really interesting argument about why NVSwitch might just be the InfiniBand of scale-up networking. A lot to get through, let's go. Skyler: Okay, first up — Nvidia and co-packaged optics. So Nvidia has been telegraphing this for a while, but now we're getting more concrete details around what they're calling their CPO strategy for the Quantum-X and Spectrum-X platforms, targeting a 2026 launch. And I want to spend some real time on this because co-packaged optics is one of those things that sounds incremental until you understand why it's actually a pretty fundamental rethink of how you get photons in and out of a switch or NIC at extreme scale. Skyler: So let's back up. The traditional model — what we'd call pluggable optics — is you've got your switch ASIC or your NIC on a PCB, and then you've got separate optical transceivers, the things that look like little metal dongles, plugged into a faceplate. That's your QSFP-DD, your OSFP world. It works. It's modular, it's proven, it's a whole ecosystem. But here's the physics problem: as you try to push bandwidth higher — we're talking 800 gig per port, 1.6 terabit per port — the electrical traces between your ASIC and those pluggable modules become a real liability. You're burning serious power just driving signal down copper at those speeds before it even becomes light. The SerDes power, the re-timers, it all adds up. Skyler: Co-packaged optics moves the optical engine — the lasers, the photodetectors, the modulator — physically much closer to the switch ASIC, either in the same package or right next to it on the same substrate. You dramatically shorten that electrical path. The result is supposed to be something like thirty to forty percent lower power for the optical I/O, which at the scale of an AI cluster is not a rounding error. We're talking potentially tens of kilowatts across a large fabric. Skyler: Now, Nvidia's angle here is that Quantum-X is their InfiniBand switch play and Spectrum-X is their Ethernet-for-A I play. Both are targeting the hyperscaler and large enterprise AI cluster market. Putting CPO on both roadmaps simultaneously tells you they're serious — this isn't a science project, it's going into product. The 2026 timeframe is tight. CPO has been "two years away" for a while in this industry, and the manufacturing and supply chain complexity around integrating photonics at package level is genuinely hard. You have to co-design the ASIC, the photonic engine, the packaging, and the thermal management all together. There's very little room to mix and match. Skyler: The competitive angle here is important too. Broadcom has been investing heavily in CPO for their Ramon and Tomahawk lines. Arista and other white-box switch players are watching closely because a Nvidia-native CPO solution further tightens the vertical integration story. If your CPO is purpose-built for NVLink and InfiniBand traffic patterns, you can optimize in ways a merchant silicon vendor targeting a broader market can't easily match. This is Nvidia trying to own more of the physics stack. And for the AI fabric world, where power and bandwidth per watt are the governing constraints, owning the physics stack matters enormously. Skyler: Alright, next story — Salience Labs and silicon photonics switching. So this one comes out of The Next Platform, and Salience Labs is a startup that's pitching a pretty ambitious vision: using silicon photonics not just for transceivers, but as the actual switching fabric. And that's a fundamentally different bet than what most of the industry is doing. Skyler: So here's the distinction that matters. Most of the silicon photonics conversation in datacenter networking is about using photonics for the last mile — getting light in and out of an electrical switch ASIC efficiently. The ASIC itself is still doing the switching, still doing the routing logic, still running SerDes. Salience is arguing that you can move the switching function itself into the optical domain. An optical switch doesn't convert to electrical, make a forwarding decision electrically, and convert back. It routes light directly. Skyler: The appeal is obvious if you're thinking about latency and power. An optical switch can theoretically be much lower latency — you're not paying for O-E-O conversion, optical to electrical to optical, at every hop. And the power story can be compelling because you're not running a massive electrical crossbar. The catch — and it's a significant catch — is that optical switches have historically struggled with the things that make Ethernet and InfiniBand fabrics useful. Buffering, for one. You can't easily buffer photons. Flow control, congestion management, multicast — all of these are hard in the optical domain. The intelligence that makes a network fabric actually work well under real AI training traffic has been built up over decades in the electrical domain. Skyler: What Salience is claiming is that they have a path to scale this up using silicon photonics manufacturing — meaning they can leverage existing CMOS fabs, which is critical for cost and scale. That's genuinely important. Custom photonic processes are expensive and hard to scale. If you can run photonic devices through a standard silicon foundry, that changes the economics dramatically. Skyler: The question I'd want answered is: where exactly does Salience think electrical intelligence lives in their architecture? Because you can't just wave away the buffering and congestion management problems for AI workloads. Training jobs generate very bursty, highly synchronized traffic — that's exactly where you need smart queuing. Either they have a hybrid architecture where optical switching handles the bulk transport and electrical logic handles edge cases, or they're betting that A I fabric traffic can be engineered to be predictable enough to sidestep those problems. Both are interesting bets. Neither is obviously wrong. But I'd want to see real workload data before I got too excited. This is one to watch — silicon photonics switching has been a research topic for a long time, but the combination of better foundry access and the specific demands of A I fabric might finally be creating the right conditions for a commercial push. Skyler: Okay, story three — power. And I mean power as in, the thing that is quietly becoming the single biggest constraint on datacenter expansion. The Next Platform ran a piece on this, and it's a topic I want to spend time on because it underpins every single other story we're covering today. Skyler: The numbers are staggering. A modern A I training cluster — we're talking something like an NVL72 rack or a comparably dense system — can pull somewhere north of one hundred kilowatts per rack. Individual GPU trays are pushing toward a kilowatt per chip. When you multiply that across a cluster with thousands of accelerators, you're looking at facilities that need hundreds of megawatts, and the largest hyperscaler deployments are now planning for gigawatt-scale campuses. A gigawatt. That's a medium-sized power plant, just for one datacenter campus. Skyler: And this hits fabric directly. More power at the chip means more heat, which changes your cooling architecture, which constrains your physical density, which affects your interconnect topology. If you're doing direct liquid cooling to the chip, your rack layout changes. If you're doing immersion cooling, your cable routing changes. Every one of those physical changes has implications for how you run fiber, how you design your top-of-rack switching, how you plan your optical paths. Fabric engineers don't get to just design in a vacuum — you're designing inside a thermal and power envelope that is getting tighter every generation. Skyler: The utility side is also becoming a real operational risk. Getting grid connections at the scale these hyperscalers need is a multi-year permitting and construction problem. We're seeing datacenters being sited specifically around power availability — nuclear plants, hydroelectric, natural gas generation. That geographic distribution of compute has network implications too, because now you're talking about low-latency interconnect across longer distances, which brings in things like coherent optics and datacenter interconnect, DCI, as a first-class fabric concern rather than an afterthought. Skyler: For the companies building power infrastructure — UPS vendors, PDU vendors, liquid cooling vendors — this is an obvious growth story. But for silicon and networking engineers, the message is: power is not someone else's problem anymore. The choices you make in your interconnect architecture have direct power consequences, and those consequences are increasingly showing up in the business case for whether a cluster gets built at all. CPO, optical switching, even the move from copper to optical at shorter reach — all of these are at least partly power-driven decisions. That's the through-line connecting basically everything happening in high-speed interconnect right now. Skyler: Story four — Vantage Data Centers raising two billion dollars. So Vantage announced a two billion dollar raise for North American early-stage portfolio development, and this is on top of what Data Center Dynamics notes is already forty billion dollars in financing they've received this year. Yes, forty billion dollars. In one year. Skyler: So Vantage is a colocation and hyperscale datacenter developer — they're not a hyperscaler themselves, they build and operate the physical facilities that hyperscalers and large enterprises lease. And the scale of capital flowing into this company tells you something really important about where infrastructure investment is right now. This isn't speculative. These are real dollars chasing real committed demand from A I buildouts. Skyler: The "early-stage portfolio development" framing is interesting. That's not just maintenance capital or expanding existing sites — that's greenfield. New land, new power agreements, new builds. And the North America focus makes sense given the current policy environment around domestic A I infrastructure. There's a lot of political and economic pressure to keep A I compute capacity onshore, and datacenter developers are positioning to capture that. Skyler: What does this mean for fabric and interconnect? A few things. First, when you're building at this scale and speed, standardization becomes incredibly important. You can't afford to have every site be a snowflake. That pushes toward standard rack designs, standard cabling plant, standard optical reach assumptions. Second, the density of these new builds — remember what we just said about power — means the physical plant decisions being made right now will lock in interconnect architectures for a decade. The fiber routing, the meet-me rooms, the intra-campus optical topology — all of that is being decided during the design phase of these greenfield builds. Skyler: And third, the sheer volume of optical components, switches, cables, and transceivers that need to fill buildings like these is putting real pressure on supply chains. We've heard a lot about transceiver shortages, fiber shortages, even structured cabling backlogs. When you've got multiple large colo developers all building simultaneously, all trying to pull from the same supply chain, component lead times stretch. That has real implications for program timelines and for which vendors have the relationships and production capacity to actually deliver. Skyler: Two billion dollars sounds like a lot. In the context of what's being built, it's almost table stakes. The datacenter infrastructure investment cycle we're in right now is unlike anything the industry has seen before, and it's directly driven by A I training and inference demand that isn't slowing down. Skyler: Alright, last story — and this one's my favorite of the week. The Next Platform published a piece arguing that NVSwitch is essentially the InfiniBand of scale-up A I networks. And I think this framing is really sharp and worth unpacking. Skyler: So let's define the terms. Scale-up networking — as opposed to scale-out — is the interconnect within a single logical compute domain. For Nvidia, the canonical example is NVLink connecting GPUs within a node or across nodes within an NVL72 or similar system. NVSwitch is the switching fabric that makes multi-GPU NVLink topologies work — it's what lets you build a non-blocking all-to-all fabric across a large number of GPUs without all-pairs direct connections. Skyler: The InfiniBand parallel is interesting because InfiniBand went through a similar evolution. It started as a general-purpose interconnect standard, got increasingly optimized for HPC and then A I training workloads, and became deeply associated with Nvidia after the Mellanox acquisition. And at some point, InfiniBand stopped being something you could really evaluate as a standalone technology independent of the Nvidia ecosystem — it became part of the integrated Nvidia compute-and-fabric stack. The argument is that NVSwitch is following the same trajectory, but for scale-up rather than scale-out. Skyler: Here's why that matters competitively. AMD, Intel, and the custom silicon players like Tenstorrent and Cerebras are all building their own scale-up interconnects. AMD has Infinity Fabric. There's U A Link, the UALink consortium, which is trying to create an open standard for scale-up interconnect that doesn't require licensing from Nvidia. The whole point of U A Link is to have a competitive alternative to NVLink and NVSwitch. But the NVLink-NVSwitch combination has compounding advantages: it's been deployed at scale, it's what CUDA and the software stack is optimized for, and every generation of H-series and B-series GPUs makes the NVSwitch fabric faster and deeper. Skyler: The InfiniBand analogy also holds in terms of vertical integration. Just as Nvidia owning Mellanox let them co-design GPUs and network adapters together, owning NVSwitch lets them co-design the GPU die and the switch fabric together. You can make architectural choices — memory access patterns, cache coherency protocols, collective operation offloads — that simply aren't available if your GPU and your switch are from different vendors with an open standard in between. That integration has real performance implications, especially for the all-reduce operations that dominate transformer training workloads. Skyler: The risk for Nvidia is the same risk InfiniBand faced — if a good-enough open alternative emerges, and if hyperscalers get motivated to build their own stacks, the moat narrows. We're seeing hyperscalers like Google with ICI, Meta with their custom fabrics, and Amazon with Trainium interconnects all pushing in that direction. U A Link is at least partly a response to not wanting to be permanently dependent on NVLink licensing. Whether any of those alternatives get to NVSwitch performance levels at scale is the question that will play out over the next few years. Right now, NVSwitch is the benchmark everything else is measured against for scale-up AI training. That's a strong position to be in. Skyler: Alright, that's the show. Nvidia bets on co-packaged optics for 2026, Salience Labs makes a swing at optical switching, power continues to reshape every datacenter decision, Vantage pulls in two billion more dollars as the build-out accelerates, and NVSwitch cements itself as the InfiniBand of scale-up fabric. A lot of threads here that are going to keep developing — I'll be watching all of them. Line Rate, brought to you by Ada and Zen Semiconductor — Ada helps companies integrate A I into their real-world workflows, go check them out. If you found this useful, subscribe wherever you get podcasts, share it with a colleague who cares about this stuff, and all the links and sources are in the show notes. See you next time.