Every transformative technology wave creates infrastructure before it creates applications. The railroad boom built tracks before it built commerce. The internet boom built fiber before it built e-commerce. The AI wave is building compute, power, and connectivity infrastructure — and we are still in the early innings.

Understanding where we are in that buildout, and which layers of the stack offer the most compelling risk-adjusted returns, is central to how Coelho Capital approaches the AI opportunity.

The Stack: Three Layers, Three Different Investment Profiles

The AI infrastructure buildout can be understood as three distinct layers, each with its own supply-demand dynamics, competitive structure, and investment characteristics.

Layer 1: Semiconductors

The chip layer is where the AI buildout is most visible and most discussed. Demand for high-performance compute — GPUs, custom ASICs, and the high-bandwidth memory that feeds them — has grown faster than the industry's ability to supply it.

The key dynamic here is not just demand growth; it is the concentration of that demand. A small number of hyperscalers account for the majority of advanced chip purchases. This creates a customer concentration risk for chip suppliers, but it also creates extraordinary pricing power in the near term, as these customers have no alternative and cannot afford to wait.

The more interesting medium-term question is the emergence of custom silicon. Several hyperscalers are now designing their own AI accelerators, reducing their dependence on merchant silicon. This is a headwind for the dominant GPU suppliers over a multi-year horizon, but it is also an opportunity for the companies providing the design tools, packaging technology, and advanced manufacturing that make custom chips possible.

We are selective in this layer. The headline names are well-owned and fully priced for a continuation of the current trajectory. We prefer exposure to the picks-and-shovels of chip manufacturing — advanced packaging, lithography, and the specialty materials that enable leading-edge nodes.

Layer 2: Data Centers

The data center layer is where the capital intensity of the AI buildout becomes most apparent. A single hyperscale AI training cluster can cost $1–2 billion to build and equip. Globally, data center construction spending is running at a pace that exceeds the prior decade's cumulative investment.

This creates opportunity across the value chain: real estate investment trusts that own and lease data center facilities, power and cooling equipment suppliers, and the fiber and networking infrastructure that connects facilities to each other and to end users.

The critical constraint in this layer is not capital — the hyperscalers have plenty of that. It is power and physical space. Permitting timelines for new grid connections in key markets (Northern Virginia, Phoenix, Singapore, Dublin) have extended to three to five years. This is creating a significant premium for existing, powered data center capacity and for the companies that have secured land and power rights in advance.

We view the data center REITs as a core holding in this theme. They offer a way to participate in the AI buildout with contracted cash flows, inflation-linked escalators, and a physical asset base that is genuinely difficult to replicate.

Layer 3: Power Grid

The power grid layer is the least-discussed and, in our view, the most underappreciated part of the AI infrastructure opportunity.

The United States electrical grid was not designed for the load growth that AI is imposing on it. Transformers, switchgear, and transmission lines that were sized for a world of stable, predictable demand are now being asked to serve data centers that consume as much power as a mid-sized city — and that need that power delivered with near-perfect reliability.

The backlog for large power transformers — the critical equipment that steps voltage up and down across the transmission system — has extended to three to four years in some cases. Domestic manufacturing capacity is limited, and the supply chain for the specialized steel and copper used in these transformers is constrained globally.

This is a classic infrastructure bottleneck: the demand is real, the need is urgent, and the supply response is slow. The companies that manufacture grid equipment, provide engineering and construction services for transmission projects, and supply the raw materials for grid buildout are positioned to benefit from a multi-year spending cycle that is largely independent of the economic cycle.

The Risk We Are Watching

The primary risk to the AI infrastructure thesis is a demand disappointment — a scenario in which the monetization of AI does not materialize at the pace that current capital spending implies. Hyperscalers are making enormous bets on AI adoption. If enterprise and consumer adoption lags, the buildout could pause or reverse, as it did in the fiber overbuild of the early 2000s.

We take this risk seriously. Our positioning in this theme is weighted toward the infrastructure layers — power, real estate, and grid equipment — that have value independent of any specific AI application. These assets serve a growing economy regardless of which AI models win. The chip layer, where valuations are most stretched and the technology cycle is fastest, receives a smaller allocation.

Our View

The AI infrastructure buildout is real, it is large, and it is durable. The question for investors is not whether to have exposure, but where in the stack to take it. We believe the power grid and data center real estate layers offer the best combination of visibility, defensibility, and valuation at this stage of the cycle.

The semiconductor layer remains important, but selectivity matters more than it did two years ago. The easy money in the headline names has been made. The next phase of returns will come from identifying the less-obvious beneficiaries of a buildout that has years left to run.

---

This commentary reflects the views of Coelho Capital Research as of the date of publication and is intended for informational purposes only. It does not constitute investment advice. Past performance is not indicative of future results.