Two AI Infrastructure Stocks for the Long Haul: Marvell and Arista

The next phase of the AI investment cycle is clear: it's moving from the software and chip layers down to the physical and electrical rails. Demand for AI compute is on an exponential S-curve, and the central constraint is no longer processing power alone-it's power. The projection is staggering: AI data center power capacity in the United States is expected to grow from about 30 or more gigawatts (GW) in 2025 to 90 or more GW or more by 2030, a compound annual growth rate of approximately 22%. That's a capacity larger than the entire power demand of California today. In practical terms, this means AI facilities are becoming equivalent to small cities in their energy needs, making power the defining operational and strategic bottleneck.

This shift is forcing a fundamental rethinking of data center economics. What was once a background infrastructure concern is now a central constraint, reshaping site selection, architectural design, and even the relationship between operators and the power grid itself. As one expert notes, 2026 marks a critical inflection point for power, with aging grid infrastructure struggling to keep pace with unprecedented load growth. The result is a market in transition, where investor rotation is already underway. The divergence in stock performance shows that investors have rotated away from AI infrastructure companies where growth in operating earnings is under pressure and capex spending is debt-funded. The message is clear: not all infrastructure plays are created equal. The durable winners will be those embedded in the high-margin, high-barrier segments of this new S-curve-the companies building the fundamental rails for the next paradigm.

Marvell Technology: The Interconnect Architect

Marvell is positioning itself as the architect of the next layer of AI infrastructure: the high-bandwidth, low-latency fabrics that connect thousands of processors across multi-rack systems. This is a critical juncture on the adoption S-curve, where the sheer scale of AI clusters demands a new class of interconnect technology. Marvell's Ultra Accelerator Link ("UALink") scale-up switch team is built for this exact challenge, designed to efficiently link large numbers of accelerators into a single, cohesive system. As AI workloads evolve from single-rack deployments to larger configurations, UALink provides the open standard fabric needed for flexible resource sharing and peak performance.

The company is aggressively expanding its switching portfolio to own this space. Its recent acquisition of XConn Technologies adds advanced PCIe and CXL switching silicon, significantly broadening its technical depth and customer reach. This move directly addresses the emerging need for memory disaggregation and high-performance connectivity in accelerated infrastructure. Marvell expects these new products to begin contributing revenue in the second half of fiscal 2027, with the acquisition projected to be accretive to earnings by fiscal 2028. More broadly, Marvell's custom ASIC designs leverage cutting-edge IP, including industry-leading 112G XSR SerDes and support for advanced process nodes like 3nm, to create silicon optimized for the unique demands of hyperscalers. This capability is key for building the fundamental rails that will carry the data of the AI era.

Arista Networks: The Power and Efficiency Layer

Arista's role in the AI infrastructure stack is defined by a relentless focus on efficiency. In a world where power is the new bottleneck, the company's new 800G R4 series portfolio sets a benchmark for low power consumption and high capacity. This isn't just incremental improvement; it's a direct assault on the total cost of ownership (TCO) for AI clusters. By delivering a dense 800 Gbps system with a 3.2 Tbps HyperPort, Arista's architecture aims to shorten AI job completion times by up to 44% compared to older methods. For hyperscalers, this translates to fewer servers needed to run the same workload, a critical lever for managing the exponential power draw of AI.

The company's current technological lead is clear. In the explosive 800GbE market, Arista led in branded market share for both 800GbE as well as overall data center Ethernet switching in the second quarter of 2025. This dominance is built on its deep software integration with the Arista EOS operating system, which ensures predictable latency and scalable protection against congestion. Yet the real test for 2026 is market share capture. The catalyst is the volume shipment of Ultra Ethernet Consortium (UEC)-compliant switches. This new standard aims to bring the simplicity and cost advantages of Ethernet to the high-performance networking traditionally dominated by InfiniBand. Arista's ability to ship these UEC-compliant systems at scale will determine whether it can leverage its current lead to displace entrenched competitors in the AI cluster market.

The bottom line is that Arista is building the efficient, high-density fabric that makes the AI power S-curve manageable. Its new R4 series directly addresses the TCO imperative, while the 2026 UEC rollout is the next major adoption hurdle. For a company positioned at the intersection of networking and power efficiency, this is the critical phase where its infrastructure layer must prove it can carry the weight of the AI paradigm shift.

Long-Term Hold Rationale and Key Catalysts

The investment case for Marvell and Arista rests on a simple, durable truth: the AI paradigm shift is a multi-year infrastructure build-out, not a fleeting software trend. While the application layer churns, the physical and electrical rails are being laid down with exponential force. For a long-term hold, the thesis hinges on two primary catalysts that will validate and accelerate this build-out.

First, the finalization and adoption of new standards in 2026 will lock in market share for the next infrastructure cycle. For Marvell, the catalyst is the custom ASIC ramp. Its 3nm designs for AWS and Microsoft are expected to hit volume production, turning design wins into cash flow. This is the critical threshold where the company's role as an IP provider for the "Build Your Own Silicon" trend must transition from promise to performance. For Arista, the catalyst is the volume shipment of Ultra Ethernet Consortium (UEC)-compliant switches. This new standard aims to bring Ethernet's simplicity to high-performance AI clusters. Arista's ability to ship these systems at scale will determine whether it can leverage its current 800GbE dominance to displace entrenched competitors, securing its position in the next generation of cluster fabrics.

Second, investors must watch for hyperscaler capex spending to continue exceeding consensus estimates. This trend has been consistently underestimated for two years, and the divergence in stock performance shows the market is becoming selective. The consensus estimate for 2026 capital spending by AI hyperscalers is now $527 billion, up from $465 billion just a quarter ago. The key is that investors are rewarding companies where this spending demonstrably links to revenue, not just debt-funded expansion. For Marvell and Arista, sustained capex from their major cloud customers is the direct fuel for their growth engines. Any sustained beat to these estimates would confirm the robustness of the AI infrastructure S-curve.

The major risk to this thesis is the pace of grid upgrades and permitting. As AI data center power capacity is projected to grow from about 30 or more gigawatts (GW) in 2025 to 90 or more GW or more by 2030, the physical grid is the ultimate bottleneck. Aging infrastructure and regulatory delays could bottleneck data center build-out despite massive capex. This creates a fundamental tension: the exponential demand for compute is outstripping the ability to deliver the power to run it. For Marvell and Arista, this risk is indirect but material. A slowdown in data center construction would delay the deployment of their advanced switches and interconnects, even if the long-term demand remains intact.

The bottom line is that both companies are positioned at the high-margin, high-barrier layers of a multi-year build-out. Their 2026 catalysts-standard adoption and capex validation-are the milestones that will prove the durability of their infrastructure plays. The grid risk is a reminder that even exponential growth has physical constraints. For a long-term hold, the focus should remain on execution against these specific milestones, while monitoring the broader power infrastructure landscape for any signs of a hardening bottleneck.

Risks and Counterpoints

The thesis for Marvell and Arista is built on a powerful, multi-year trend. Yet, in the real world of technology adoption, even the steepest S-curves face friction. The specific challenges for each company highlight the gap between design wins and durable market dominance.

For Marvell, the primary risk is execution. The company is attempting a rapid integration of new capabilities, most notably the acquisition of XConn Technologies to bolster its PCIe and CXL switching portfolio. While the talent and IP are compelling, merging engineering teams and product lines under pressure is a classic integration risk. More fundamentally, Marvell's entire growth story rides on its role as an IP provider for the "Build Your Own Silicon" trend. This exposes it to a powerful counterforce: the hyperscalers themselves. As custom ASIC projects for AWS and Microsoft are expected to hit volume production, the risk is that these customers will eventually internalize more of the design work, particularly for the high-speed SerDes blocks Marvell provides. The company must maintain its technological lead and secure long-term design wins to avoid being squeezed out of its own growth engine.

Arista faces a different kind of adoption risk. Its 2026 catalyst is the volume shipment of Ultra Ethernet Consortium (UEC)-compliant switches. This new standard is a direct challenge to the entrenched InfiniBand ecosystem and proprietary solutions from vendors like Nvidia. The pace of adoption is uncertain. If the UEC standard fails to gain critical mass, or if competing solutions from hyperscalers or other vendors prove more compelling, Arista's current 800GbE market share lead may not translate into the next generation of cluster fabrics. The company's software integration with EOS is a moat, but it must be strong enough to lock in customers during this pivotal standard-setting phase.

A broader, systemic risk looms over both companies: the physical limits of power. The exponential growth in AI demand is projected to push U.S. data center power capacity from about 30 or more gigawatts (GW) in 2025 to 90 or more GW or more by 2030. This is a staggering build-out, but it is entirely dependent on the grid's ability to keep pace. Regulatory limits on emissions, permitting delays for new power lines, or the sheer cost of upgrading aging infrastructure could cap the total addressable market for AI data centers. In this scenario, even the most advanced switches and interconnects would have fewer racks to fill. This is the ultimate bottleneck, a physical constraint that could moderate the entire S-curve, regardless of technological prowess.

The bottom line is that the path to exponential growth is paved with execution hurdles and competitive battles. For Marvell, it's about integrating acquisitions and defending its IP role against in-house designs. For Arista, it's about winning the standard-setting race in a crowded field. And for both, the entire paradigm shift is contingent on a power grid that can scale as fast as the demand it must serve.