7 AI Infrastructure Stocks Beyond the Chips - The Motley Fool

Skyler: Welcome to Line Rate — the show covering datacenter fabric, A I interconnect, and the hardware moving data at the speed that modern A I demands. I'm Skyler Monroe, and if you care about what's happening inside and between the racks, you're in the right place. Big thanks to our sponsors: Ada, which helps businesses 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 — including Ada for A I deployment at scale. Today we've got a packed show: Arista's transformation into an A I networking powerhouse, what the GenAI boom is doing to Ethernet switching, Marvell going deeper on silicon germanium with GlobalFoundries for optical applications, and we'll touch on the broader investment landscape around A I infrastructure. Let's get into it. Skyler: Alright, let's start with Arista Networks, because this one has been brewing for a while and it's hitting a tipping point. There's a piece out from Dr. Robert Castellano's Semiconductor Deep Dive newsletter making the case that Arista has genuinely morphed — his word, morphed — from a datacenter networking vendor into an A I networking powerhouse. And look, that framing is doing a lot of work, so let's unpack what's actually happening technically and competitively. Skyler: For years, Arista's identity was built on EOS — their Extensible Operating System — and high-density Ethernet switching. They were the go-to for hyperscaler front-end networks. Big leaf-spine deployments, massive route tables, telemetry that actually worked. That was the business. But front-end networking, as clean and profitable as it is, is not where the growth story is right now. The growth story is the A I back-end fabric — the GPU-to-GPU, accelerator-to-accelerator traffic that looks completely different from traditional east-west flows. Skyler: A I back-end traffic is bursty, it's latency-sensitive in a very specific way, and the failure modes are brutal. If you lose a flow in a traditional application, you maybe retry. In a distributed training job across thousands of GPUs, a stalled flow can stall the entire job. So the requirements on the switching layer are genuinely different — you need ultra-low and consistent latency, you need deep buffers in the right places, and you need the control plane to be smart about load balancing across a fabric that's doing collective communications, not just unicast flows. Skyler: Arista's play here has been pushing hard into this space with their 7800 series and the broader A I spine architecture. They've been positioning around adaptive routing — their ECMP and load balancing capabilities — and leaning into the fact that their software stack is battle-tested at hyperscaler scale. The pitch is essentially: you already trust Arista for your front-end, why introduce a new vendor and a new operational model for the back-end? Skyler: Now the competitive context here matters a lot. Arista is going up against Nvidia's InfiniBand ecosystem on one side, and on the other, they're actually a potential beneficiary of the Ultra Ethernet Consortium — U E C — gaining traction. U E C is trying to define Ethernet that's actually fit for A I workloads, with congestion control, latency guarantees, and multipath forwarding that InfiniBand users have taken for granted. Arista is a member of U E C, and if Ethernet-based A I fabrics become the dominant model at scale — which is far from settled but increasingly plausible — Arista is well positioned. Skyler: The caution I'd add is that this transformation isn't complete. A I networking revenue is growing fast for Arista, but it's still a company whose roots and most of its current book is traditional datacenter Ethernet. The question is whether they can execute the pivot deeply enough before more specialized competitors — whether that's Nvidia on the InfiniBand side or startups building A I-native silicon — carve out the high-value A I fabric territory. It's a real race and Arista is running hard. Worth watching closely. Skyler: Next up — The Next Platform had a piece on how the GenAI boom is accelerating and transforming Ethernet switching, and I want to spend some time here because this is where the rubber meets the road for a lot of the standards conversations we follow on this show. Skyler: The core argument is that GenAI isn't just driving more traffic — it's changing the fundamental character of the traffic in ways that stress Ethernet switching architectures that were designed for a different world. Traditional datacenter Ethernet was optimized for web-scale workloads: lots of flows, mix of large and small, relatively tolerant of jitter, with TCP doing its thing in the background to manage congestion. The switching silicon was built around that model. Skyler: GenAI training workloads look nothing like that. You've got all-reduce and all-to-all collective communications patterns where every node needs to talk to every other node essentially simultaneously. The traffic pattern is highly synchronized — think of it as a wave — and when that wave hits a switch fabric that's doing traditional ECMP, you get congestion hotspots that tank your training throughput. The GPU utilization numbers can fall off a cliff. Skyler: So what's the industry actually doing about this? A few things happening in parallel. First, there's a push for smarter adaptive routing at the switch level — not just static hash-based ECMP but dynamic per-packet or per-flow load balancing that reacts to real-time congestion. Broadcom's latest generations of switching silicon have been building this in. Cisco, Arista, and others are implementing it at the platform level. But the software complexity of doing this right, at scale, without creating new failure modes, is non-trivial. Skyler: Second, there's the buffer story. A I training fabrics need different buffer architectures than traditional switching. You want enough buffer to absorb the synchronized bursts without tail-dropping, but not so much that you're introducing latency. The balance point is different from what years of web-scale optimization landed on, and it's driving real silicon architectural changes in the merchant silicon ecosystem. Skyler: Third — and this is where it connects to everything else we talk about on Line Rate — there's the question of whether Ethernet can close the gap with InfiniBand on the A I-specific features that actually matter. U E C is the standards effort trying to answer that. The spec is still evolving — I want to be clear about that, this is not a done deal — but the direction is toward a profile of Ethernet with defined congestion control behavior, latency targets, and multipath semantics that are actually suitable for collective communications. Skyler: The transformation the Next Platform is pointing to is real. The Ethernet switching market is not just growing because there are more GPUs — it's being fundamentally redesigned around A I fabric requirements. And that's creating both opportunity and disruption across the switching silicon and systems stack. Companies that were optimized for the old model are having to adapt fast. Skyler: Okay, let's talk optics. Marvell is expanding its silicon germanium — SiGe — capacity through a partnership with GlobalFoundries, specifically aimed at A I datacenter optical applications. This one caught my attention because it's a manufacturing strategy story as much as a technology story. Skyler: First, a quick grounding on why silicon germanium matters here. SiGe is a process technology that enables very high-speed analog and mixed-signal circuits — we're talking the kind of circuitry you need for optical transceivers operating at 100 gigabaud and beyond per lane. For co-packaged optics and high-speed pluggable modules, the DSP and driver circuitry that interfaces between the electrical and optical domains benefits enormously from SiGe's performance at high frequencies. It's not the only approach — you can also do this in standard CMOS with enough engineering — but SiGe has real advantages in certain high-frequency performance metrics. Skyler: Marvell has been a significant player in optical DSP silicon — their Colorz and PAM4 DSP lines, and their broader photonics strategy. Expanding SiGe capacity with GlobalFoundries signals a few things. One, they're betting on sustained and growing demand for high-performance optical front-ends as the datacenter transitions to higher line rates. The move to 800 gig and eventually 1.6 terabit per second optical modules is not hypothetical — hyperscalers are qualifying these now or will be shortly. Skyler: Two, this is a supply chain diversification play. The optical component market has had capacity constraints, and having dedicated capacity locked in with a foundry partner like GlobalFoundries gives Marvell more control over their supply and potentially their cost structure. For a component that's increasingly on the critical path for A I cluster builds, that's not a trivial advantage. Skyler: Three — and this is the angle I find most interesting from a datacenter fabric perspective — this move is happening in the context of the co-packaged optics debate. The industry has been wrestling with whether pluggable optics or co-packaged optics win in the A I fabric context. Co-packaged optics, where the optical engines are integrated much more closely with the switch ASIC, promises better power efficiency and higher bandwidth density. But it's a massive supply chain and ecosystem disruption. Pluggables are winning in the near term because the ecosystem exists. Skyler: Marvell expanding SiGe capacity suggests they're positioning for both paths — high-performance pluggables now, and the optical engine silicon that will be needed for co-packaged solutions as that architecture matures. The technology foundation is the same even if the packaging is different. Watching Marvell's moves here tells you something about where the serious money thinks optical interconnect is going over the next three to five years. Skyler: Now let me spend a few minutes on the broader investment picture, because there were a couple of pieces — one from Motley Fool, one touching on the photonics space through the lens of a Jim Cramer call — that are worth addressing even if they're investor-facing rather than purely technical. Skyler: The Motley Fool piece was framing out seven A I infrastructure stocks beyond the chips — essentially the argument that Nvidia gets all the oxygen in the room but there's a broader ecosystem of companies building the infrastructure that makes A I work at scale. That ecosystem includes networking, optics, power distribution, cooling, and software. And look, from where I sit, this framing is absolutely correct technically. The GPU is not the whole story. The fabric is the story. Skyler: The reason I bring this up on Line Rate isn't to give you stock tips — that's not what we do here — but because the investment attention flowing into this space is actually a signal about where the engineering problems are hard and where capital is going to drive solution development. When public market investors start caring about A I networking and photonics as distinct investment theses, it means procurement budgets at hyperscalers and large enterprises are real, it means the build-out is sustained, and it means the companies solving these problems have room to invest in R&D at a scale that changes the standards landscape. Skyler: The photonics angle in particular — the discussion around companies in the optical transceiver and photonic integration space — is worth watching because that's an area where the technical differentiation is deep, the capital requirements are high, and the market is consolidating around a smaller number of serious players. Whether it's the established players in the pluggable market or newer entrants trying to crack co-packaged optics, the companies that win here are going to shape what the A I fabric looks like at the physical layer for the next decade. Skyler: I'll just say for context on the Cramer mention — the specific stock picks are outside our lane — but the underlying thesis that photonics for A I datacenters is a significant growth area is well-supported by everything happening in the standards bodies and at the hyperscalers right now. The infrastructure build is real, the optical component demand is real, and the technology is advancing fast. Skyler: Alright, let's land the plane. Big week for A I fabric and networking. Arista is making a genuine play to own the A I back-end switching narrative, not just maintain their front-end dominance — and the U E C ecosystem gives them tailwind if Ethernet-based A I fabrics win. The GenAI boom is doing real violence to traditional Ethernet switching architectures and forcing silicon and systems vendors to adapt fast — smarter routing, different buffer architectures, and new performance profiles. Marvell is locking in SiGe capacity with GlobalFoundries, which is a serious bet on the trajectory of high-speed optical interconnect for both pluggable and co-packaged futures. And the broader investment attention flowing into A I infrastructure — networking, optics, power, cooling — is a signal that the build-out is sustained and the hard problems are getting funded. Skyler: Thanks for listening to Line Rate. The show is brought to you by Ada — which helps companies actually integrate A I into their workflows and make it stick — and Zen Semiconductor, an umbrella company with a portfolio across the silicon industry. If you got value from today's episode, please subscribe wherever you get your podcasts, share it with a colleague who lives in the datacenter stack, and check the show notes for links to all the stories we covered. See you next time.