Nvidia, Constellation, Vertiv: Five Companies Collecting AI Infrastructure Rent

Published by James Harris on June 24, 2026June 24, 2026

Five companies sitting underneath every AI model are collecting infrastructure rent regardless of which lab wins the race. With the five largest US hyperscalers projected to spend over $650 billion on AI infrastructure in 2026, nearly double the roughly $380 billion deployed in 2025, the physical stack powering that buildout has quietly become its own investable category.

The logic is straightforward: every model, whether GPT, Gemini, or Llama, runs on the same five layers. Chips handle computation, servers rack the chips, power keeps them online continuously, cooling prevents thermal failure, and networking binds thousands of chips into a single coherent machine. Nvidia holds 85 to 90 percent of AI data center compute and its CUDA ecosystem keeps switching costs high enough that the position is structurally sticky. Constellation Energy anchors the power layer with the largest US nuclear fleet and a 20-year Microsoft supply deal. Vertiv manages cooling as air systems become obsolete at rack densities above 100 kilowatts, with order growth of 252 percent and a co-design relationship with Nvidia. Arista Networks reported $9 billion in fiscal 2025 revenue with Meta and Microsoft as anchor customers. The industrial-scale power procurement that Bitcoin miners dominated after 2017 has effectively been absorbed by this infrastructure tier, which arrives with sovereign-grade capital access miners never had.

None of these five layers can be skipped, and none of them require a prediction about which AI lab survives the next 18 months.

That framing matters more than it sounds. Anthropic’s reported valuation trajectory and the competitive pressure from Mistral and others mean the model layer remains genuinely uncertain. Infrastructure companies do not carry that uncertainty. The regulatory picture adds another dimension: federal involvement in permitting disputes over AI power plants signals that government agencies are beginning to treat AI energy infrastructure as a category requiring direct oversight, which could create uneven compliance costs across the stack and favor larger, better-capitalized operators. Capital concentration at this layer also connects to a broader institutional rotation narrative, where long-duration infrastructure bets are absorbing pools of capital that might otherwise have cycled through risk assets.

Super Micro rounds out the five as the fastest Nvidia deployment partner for liquid-cooled rack-scale systems, with a market cap of approximately $17 billion against the backdrop of the broader buildout. The gap between that figure and Nvidia’s roughly $5.1 trillion valuation illustrates how unevenly the market has priced the five layers so far, which is either an inefficiency or a reflection of genuine moat differences depending on which layer you examine.