AI Infrastructure: Why Buildout Matters More Than Apps

Key Takeaways:

• AI may be approaching an inflection point in its buildout. We see parallels to the early internet and electrification eras, when infrastructure buildouts followed initial hype cycles.
• Scaling AI requires physical inputs. Compute, energy, data centers, and automation are the binding constraints on AI growth.
• Infrastructure may capture more durable value. Scarcity, pricing power, and high barriers favor the physical layers over applications.

Is the Market Still Thinking About AI Like a Software Story?

Most investors still reach for AI exposure through software and application companies. That framing made sense in the early hype phase, but it may be missing where the durable value is building now. The more useful parallel is the internet buildout of the late 1990s, where the companies laying fiber and building data centers generated more durable returns than most of the apps built on top of them.

We are in a similar transition today. The AI application layer is real and growing, but it runs entirely on physical infrastructure. The companies that build and supply that infrastructure are the ones facing structural demand that does not depend on which AI application wins.

Why Is AI an Industrial System, Not Just a Software Layer?

AI at scale is a system of interdependent physical and digital components that must all expand together. Training a large model can require as many as 100,000 or more chips running in parallel for the largest models, connected by high-speed networking, cooled by industrial systems, and powered by reliable electricity. Inference, the process of running the model to generate outputs, multiplies those requirements across every user and every query.

Each of those components is a physical bottleneck. You cannot train faster models by writing better code alone. You need more and better chips, more power, more cooling, and more data center space. That is an industrial problem, and in our view it is generating industrial-scale demand.

Is the AI CapEx Cycle Just Beginning?

The first wave of AI infrastructure spending focused on compute and data centers. That wave is not over, but it is expanding. The next phase is pulling in energy infrastructure, power management, networking equipment, and physical automation. Hyperscalers have announced combined capital expenditure plans approaching $400 billion for 2025 alone (source: company filings), and most of that spending flows directly into physical infrastructure.

The important point for investors is duration. This is not a single-year capex event. Large infrastructure projects take years to plan, permit, and build. We view the demand signal from AI as long-dated and relatively visible compared with prior technology cycles.

Why May Infrastructure Capture More Durable Value Than Applications?

The economics of infrastructure and applications are structurally different. Infrastructure benefits from scarcity, pricing power, and high barriers to entry. Applications face competition, commoditization, and rapid product cycles.

What Are the Key Pillars of the AI Infrastructure Stack?

Five segments make up the investable AI infrastructure stack:

1. Semiconductors are the foundation: every AI workload runs on chips, and advanced chip design is concentrated among a small number of companies.
2. Data centers are the physical home of AI compute, and demand for new capacity is running well ahead of supply.
3. Energy is the binding constraint that most investors underestimate: a single large AI data center can require hundreds of megawatts of power — comparable to a small city’s load.
4. Industrial automation is both an input to AI buildout and an output of it, as AI accelerates the adoption of robots and automated systems in manufacturing.
5. Networking connects all of it, and the bandwidth requirements of modern AI clusters are driving a new generation of high-speed interconnect investment.

Are Supply Constraints Signals of Structural Demand?

Power shortages, chip supply constraints, land availability, and cooling limitations are regularly cited as risks to AI infrastructure growth. While these are real risks, persistent constraints despite record capex may also signal that demand is outpacing supply — a condition that historically has supported pricing power for companies at the chokepoints.

Companies operating at constrained chokepoints in the AI infrastructure stack, whether in advanced packaging, power management, data center cooling, or transmission infrastructure, may be beneficiaries rather than victims of these constraints.

How Can Investors Access AI Infrastructure Exposure?

The VanEck Semiconductor ETF (SMH) offers concentrated exposure to the chip companies at the core of the AI infrastructure buildout. For investors seeking exposure specifically to fabless semiconductor companies, the VanEck Fabless Semiconductor ETF (SMHX) focuses on chip designers that outsource manufacturing, a segment that includes many of the chip designers driving advanced AI compute architectures. Together, SMH and SMHX offer complementary ways to access the semiconductor layer of the AI infrastructure stack.