AI Infrastructure: The Picks & Shovels Play

The AI infrastructure buildout represents one of the largest capital expenditure cycles in technology history. Hyperscaler AI capex crossed $200B in 2025 and shows no signs of slowing. But while attention focuses on foundation model companies, the more compelling risk/reward may sit one layer down — in the infrastructure providers who sell the tools.

The Semiconductor Layer

NVIDIA dominates AI training compute with an estimated 80%+ market share in data center GPUs. But the competitive landscape is shifting as AMD gains traction with its MI300 series, Broadcom expands its custom ASIC business, and hyperscalers invest in proprietary silicon.

The multiple compression across the group tells an important story. Despite revenue growth rates that would be extraordinary in any other sector, the market is beginning to price in cyclicality risk. The question for investors: is this a buying opportunity or an appropriate repricing?

The Networking Layer

Every AI cluster needs high-bandwidth networking. Arista Networks has emerged as the dominant provider of data center switches for AI workloads, with its 400G and 800G Ethernet solutions displacing traditional InfiniBand in many deployments.

The networking TAM for AI infrastructure is often overlooked. Each GPU cluster requires switching capacity that scales with the number of GPUs, creating a multiplier effect on networking revenue as AI clusters grow larger.

Cloud Infrastructure: The Distribution Layer

Amazon Web Services, Microsoft Azure, and Google Cloud collectively account for ~65% of global cloud infrastructure spending. Their AI strategies differ meaningfully:

• AWS leads on custom silicon (Trainium, Inferentia) and breadth of services
• Azure benefits from the OpenAI partnership and enterprise distribution
• Google Cloud leverages TPU advantages and Gemini integration

Valuation Framework

The key question for infrastructure investors is how to think about valuation when growth rates inevitably normalize. We present a range-based framework across methodologies:

The Investment Case

The bull case for AI infrastructure rests on three pillars: (1) training compute demand is growing faster than Moore's Law can reduce costs, (2) inference demand will eventually dwarf training as AI models deploy at scale, and (3) the infrastructure layer captures value regardless of which AI models win.

The bear case centers on cyclicality, customer concentration (hyperscalers represent 50%+ of revenue for most infrastructure names), and the emergence of more efficient architectures that reduce compute requirements per unit of AI output.

Key Takeaways

For investors looking to gain AI exposure with better risk-adjusted returns, the infrastructure layer deserves serious consideration. The companies profiled here generate real cash flows, maintain high margins, and benefit from the AI buildout regardless of which foundation model ultimately dominates.

The picks and shovels thesis has worked before — from cloud computing to mobile internet. The question is whether AI infrastructure follows the same playbook, or whether this cycle's unprecedented scale creates different dynamics. We think the weight of evidence favors the former.