AI Infrastructure Enters Its Largest Architectural Transition Cycle: Bottlenecks Become Visible, Creating New Sources of Alpha

The most important development in AI this year is not simply accelerating demand growth, but the simultaneous expansion of demand and structural changes across the underlying infrastructure stack.

On one side, AI capex continues to expand rapidly and is increasingly moving toward a trillion-dollar scale. On the other, data center infrastructure is entering one of its fastest periods of change in decades. As the industry moves from GB systems toward Rubin, Rubin Ultra, and eventually Feynman architectures, the shift extends far beyond GPUs themselves — affecting the entire system stack.

Higher power density, larger clusters, and rapidly growing inference demand are driving synchronized upgrades across racks, power infrastructure, networking, storage, and interconnect technologies.

More importantly, this transition is moving from roadmap discussions into deployment and delivery. Many infrastructure changes previously remained largely theoretical — embedded in product plans, technology roadmaps, and demand expectations. As AI capex increasingly translates into physical deployment, system constraints are becoming visible in real-world implementation.

Recent disclosures increasingly support this view. Dell continues to raise AI server guidance, while NVIDIA’s networking business maintains strong growth. In our view, these developments suggest AI infrastructure spending is evolving from isolated compute procurement toward system-wide expansion.

Memory Remains One of the Most Visible Constraints

Memory demand continues to increase as longer context windows, agentic workflows, and inference workloads expand. Supply improvements, however, remain relatively gradual.

HBM, DRAM, and NAND are becoming increasingly critical components in determining effective compute utilization. We continue to believe memory supply constraints are unlikely to ease materially before 2027, making storage-related bottlenecks one of the more visible constraints across the AI stack.

Connectivity Becomes Increasingly Important for System Efficiency

As system complexity increases, network efficiency is becoming central to lowering total system costs and improving compute utilization.

Growing rack bandwidth requirements, larger clusters, and rising power density are driving networking architecture changes. Inside the rack, higher connection density is supporting migration toward NPO/CPO architectures, more complex switching topologies, and new interconnect protocols.

At the same time, the rapid growth of inference workloads is creating more distributed deployment patterns, expanding bandwidth demand beyond scale-up architectures into scale-out networking and DCI infrastructure.

Industry commentary increasingly validates this transition. Marvell continues highlighting growth opportunities in 800G/1.6T optical connectivity, switching silicon, NPO/CPO architectures, DCI, custom XPU platforms, and XPU attach opportunities. Nokia reported strong growth in AI and cloud-related demand while noting that many design wins have yet to convert into revenue. Cisco similarly continues emphasizing accelerating hyperscaler demand for networking and optical infrastructure.

Collectively, these signals suggest connectivity is increasingly moving from a theoretical investment theme toward an active deployment cycle.

Power Infrastructure Is Emerging as a System Constraint

Power systems are increasingly shifting from a supporting function toward a core system constraint.

As rack power continues rising, traditional low-voltage, multi-stage conversion architectures face growing challenges from conversion losses, thermal constraints, copper intensity, and space limitations.

The transition is not simply about moving from 48V toward 800V architectures. Rather, it represents a broader redesign of the power delivery stack: reducing conversion stages, increasing voltage levels, and moving power distribution closer to the load in order to improve efficiency and deployment density.

Technologies including 800V HVDC, ±400V architectures, high-voltage devices, BBU/CBU systems, and next-generation power distribution schemes are all increasingly important because they address constraints created by higher power density.

Bottlenecks Continue Migrating Upstream

As system complexity rises, bottlenecks increasingly migrate upstream.

Power semiconductors, optical components, advanced materials, thermal materials, and critical substrates — previously viewed largely as enabling technologies — are increasingly becoming determinants of deployment speed and system efficiency.

This transition is already becoming visible across portions of the supply chain. As demand accelerates faster than industry capacity expansion, selected components have begun experiencing pricing pressure, including MLCCs, power devices, and indium phosphide-related products.

This suggests bottlenecks are unlikely to remain confined to racks, networking equipment, or power systems. Instead, constraints increasingly propagate toward materials, devices, and manufacturing capabilities.

Overall, we believe the defining characteristic of today’s AI cycle is not simply larger investment volumes, but the combination of accelerating deployment, rapid architectural shifts, and increasingly visible system constraints.

As growth, architecture transitions, and deployment occur simultaneously, bottlenecks become more visible — and industry profit pools increasingly migrate toward whichever constraints become most binding.

In previous years, much of the discussion focused on long-term technology roadmaps. Today, we believe the industry is increasingly entering an execution and delivery phase, where identifying evolving bottlenecks becomes increasingly important for generating excess returns.

Portfolio Positioning: Maintaining Structural Exposure While Increasing Focus on Fundamental Validation

Following the recent rally, valuations across many AI-related assets are no longer inexpensive.

However, considering accelerating commercialization, continued capex expansion, and improving profitability resulting from architectural shifts, we believe valuations remain broadly explainable by fundamentals.

Our investment framework remains focused on structural opportunities, but increasingly emphasizes fundamental validation and dynamic pricing discipline.

We remain focused on memory, networking and interconnects, power infrastructure, and upstream components and materials where system constraints appear most visible. At the same time, we continue tracking deployment progress and monitoring where the next bottlenecks may emerge.

At the portfolio level, we expect positioning to remain dynamic, adjusting exposure as deployment visibility, industry pricing, and supply-demand conditions evolve.

We believe that during periods of rapid architectural change and accelerated deployment, the evolution of system constraints will continue to shape both structural opportunities and sources of excess returns.