AI data center plumbing just stole the spotlight. Cisco dropped a new 102.4 Tbps switch silicon and liquid-cooled systems at Cisco Live EMEA, and the market did what it always does when the capex gods get loud: it rotated into the picks-and-shovels that actually move bits. While mega-cap tech set the backdrop with Microsoft firm, Apple soft, and Amazon mixed, the action clustered in the AI networking complex, where Ethernet is back on offense and optics are the new oil.
What drove attention today: Cisco unveiled the Silicon One G300, a 102.4 Tbps switch chip to power new N9000 and 8000 systems, plus 1.6T optics and liquid-cooled designs. The pitch is simple and aggressive: higher bandwidth density, nearly 70 percent energy efficiency improvement on the liquid-cooled systems, and a management layer via Nexus One to scale AI fabrics with less pain. Cisco also threw out claims of 33 percent higher network utilization and a 28 percent cut in job completion time with Intelligent Collective Networking versus non-optimized paths. Quick trading profile: Dividend tech with balance-sheet depth, a valuation discount to pure-play data center winners, and a long history of selling boxes to the biggest buyers on earth. AI optionality just got less theoretical. Key takeaway: Cisco is trying to turn Ethernet into the default AI backend at scale. If even a slice of hyperscaler and sovereign AI builds move away from closed fabrics, CSCO’s margin mix and narrative get a tailwind. Track design-win disclosures and optics attach more than headlines.
What drove attention today: Cisco’s launch validated what Arista’s been evangelizing for a year: Ethernet is ready to run AI clusters, not just web search. With broad 400G and 800G portfolios and a roadmap to 1.6T, Arista sits in the slipstream of every hyperscaler capex budget that tilts toward open, deterministic Ethernet for training and inference. Customer concentration with giants like Microsoft and Meta keeps it front-row in AI fabric decisions. Quick trading profile: High-margin software-like model wrapped around hardware, net cash, premium multiple, and a habit of beating and raising when cloud customers are in harvest mode. It trades like AI networking beta, because it is. Key takeaway: If Ethernet keeps taking workloads from InfiniBand and if AI inference sprawl accelerates across enterprises, Arista stays a core way to own the pipes. Watch 800G to 1.6T transition timing, pricing discipline if Cisco turns up the heat, and any new footprint disclosures tied to AI pods.
What drove attention today: Cisco spotlighted 1.6T OSFP optics and liquid cooling, and that’s catnip for Broadcom’s portfolio. AVGO sells the switch silicon that powers a big chunk of Ethernet fabrics, PAM4 gear that enables 800G and 1.6T links, and is deep in co-packaged optics and custom accelerators for hyperscalers. When the market hears “more bandwidth per rack, less power per bit,” it hears Broadcom’s order books getting thicker. Quick trading profile: Mega-cap compounder with enviable margins, disciplined M&A, and direct leverage to the AI capex cycle across networking, storage, and compute offload. The multiple rests on the belief that AI infrastructure demand is less a spike and more a slope. Key takeaway: AVGO is a volume beneficiary if Ethernet wins share and if optics mix keeps stepping up. The risk isn’t demand; it’s bottlenecks and price elasticity in the optics chain. Keep an eye on lead times, hyperscaler capex guides, and how fast 1.6T modules go from demo to deployment.
What drove attention today: Every AI scale-out story runs through optical links, and that runs through PAM4 DSPs and coherent tech where Marvell is a key supplier. Cisco leaned hard into 1.6T OSFP and power efficiency, both of which put MRVL’s silicon in the conversation for switch-to-NIC and switch-to-server links at AI clusters. As inference proliferates beyond hyperscalers, the addressable market widens from a few buyers to many. Quick trading profile: Mid-cap with clear AI exposure, hyperscaler demand that can be lumpy quarter to quarter, and gross margins on an upswing as newer nodes and custom silicon scale. It trades on the slope of bandwidth growth and the cadence of design wins. Key takeaway: MRVL remains a clean way to own the optical bandwidth upgrade cycle. The bull case is simple: 800G today, 1.6T tomorrow, and more lanes everywhere. Execution on 1.6T DSPs, share versus competitors, and visibility into custom compute and DPU programs will drive the next leg.
What drove attention today: Cisco’s pitch sharpens the core debate of 2026 infrastructure: Ethernet AI fabrics versus Nvidia’s InfiniBand-led, tightly integrated stacks. Nvidia isn’t ignoring Ethernet; its Spectrum-X line targets AI Ethernet with congestion control, and its NVLink and InfiniBand still dominate the largest training clusters. Any narrative that shifts bandwidth and control away from Nvidia is going to get oxygen on days like this. Quick trading profile: The market’s center of gravity for AI compute with a towering valuation built on sustained demand, software lock-in, and supply chain execution. Volatility is a feature, not a bug, as investors handicap architecture transitions and who controls the data plane. Key takeaway: Even if Ethernet takes more share, Nvidia’s grip on high-performance training won’t evaporate. The risk is less binary and more about profit pool split. If open Ethernet fabrics become default for a larger slice of inference and mid-scale training, Nvidia’s networking attach gets pinched. If not, the moat gets deeper.
This sector move is about who captures the AI dollar beyond the GPU. Cisco’s G300 claims—33 percent higher network utilization, 28 percent faster completion times, and 100 percent liquid-cooled systems with big energy savings—speak to the new constraint in AI: moving data without lighting the power bill on fire. For hyperscalers, neoclouds, and sovereign builds, this is budget math, not marketing fluff. If Ethernet can consistently deliver deterministic performance at scale with simpler ops via tools like Nexus One, procurement gets easier, not harder.
Expect knock-on action beyond the obvious five. Server assemblers tied to liquid cooling, fiber and cable suppliers feeding 800G and 1.6T runs, and test and measurement names that certify these links all ride the same wave. Meanwhile, the macro canvas is cooperative enough: Microsoft strength underscores cloud budgets, Alphabet is steady, Amazon mixed but spending where it counts, and Apple doing its own gravity check in devices. None of that changes the fact that AI capex remains the single most durable line item in tech.
AI networking is where the next round of ROI gets unlocked. The GPU wars grabbed the headlines, but bandwidth, optics, cooling, and deterministic fabrics will decide who turns AI spend into profit. Own the names with real silicon, real shipments, and real hyperscaler seat time, and avoid the tourist traps that only rally on buzzwords.
