Key Points

  • GPUs get the headlines, but a data center can't compute without power delivery, cooling, high-speed networking, and memory — each a distinct bottleneck with its own set of suppliers.
  • Vertiv (NYSE: VRT) and Eaton (NYSE: ETN) sit on the physical layer of power and thermal management; Arista Networks (NYSE: ANET) owns the switching fabric; Micron Technology (NASDAQ: MU) supplies the high-bandwidth memory that feeds accelerators.
  • Bottleneck suppliers tend to hold pricing power because their products are mission-critical, hard to substitute, and often sold on long design cycles.
  • The trade-off: these businesses are cyclical, capital-intensive, and exposed to the same spending swings that lift them in an upturn.

When investors talk about the artificial intelligence buildout, the conversation usually starts and ends with graphics processing units. But a rack of the world's fastest accelerators is inert without electricity to run it, a way to keep it from overheating, a network to move data between chips, and memory fast enough to keep those chips fed. Each of those functions is a potential bottleneck — and in a supply-constrained boom, the companies that control the bottleneck often capture outsized economics.

Here are four suppliers that operate one layer removed from the marquee chipmakers, and the reasons their positions can translate into durable pricing power.

Power: The First Constraint Nobody Priced In

Modern AI training clusters draw enormous amounts of electricity, and the density of power packed into a single rack has climbed sharply as accelerators have grown more capable. That shift turns power distribution from a background utility concern into a first-order engineering problem.

Eaton (NYSE: ETN) operates in electrical power management — the switchgear, distribution equipment, and backup systems that route and protect electricity inside a facility. Data centers are only one end market for Eaton, which also serves aerospace, industrial, and utility customers, but the surge in data center construction has made electrical infrastructure a focal point of demand.

The pricing-power argument here is structural. Grid interconnections and permitting can take years, high-capacity electrical gear has long lead times, and a power failure in a live AI cluster is catastrophically expensive. When customers are racing to bring capacity online, the scarce input — reliable, high-density power equipment — commands a premium, and suppliers with established manufacturing footprints are hard to route around.

Cooling: Where Heat Becomes the Bottleneck

Every watt of power that enters a chip leaves as heat. As rack densities climb, traditional air cooling reaches its physical limits, pushing operators toward liquid cooling and more sophisticated thermal management.

Vertiv (NYSE: VRT) specializes in this layer, providing power and thermal management systems purpose-built for data centers, including liquid-cooling technology designed for high-density deployments. Cooling is not a commodity add-on; it has to be engineered around a specific rack architecture, integrated with power delivery, and serviced over the life of the facility.

That integration is the moat. Once a hyperscaler or colocation operator standardizes on a cooling architecture, switching costs rise, and the supplier that helped design the system is well positioned for follow-on orders and service revenue. Vertiv's exposure is more concentrated in critical digital infrastructure than a diversified industrial like Eaton, which cuts both ways — more leverage to the AI theme, and more sensitivity to any pause in it.

Networking: Stitching Thousands of Chips Together

An AI cluster is not one giant computer; it is thousands of accelerators lashed together and forced to behave as a single machine. The speed and efficiency of the network connecting them directly determines how fast a model trains. Networking is therefore not plumbing — it is a performance bottleneck in its own right.

Arista Networks (NYSE: ANET) builds high-performance Ethernet switching and the software that runs it, and counts large cloud operators among its most important customers. As the industry pushes to scale AI back-end networks, the debate over which networking standards win is central to how these clusters get built, and Arista is a prominent player in the Ethernet camp.

Pricing power in networking comes from a combination of raw performance and software stickiness. Operators standardize on an operating system and management tooling, train their engineers on it, and build automation around it — all of which raise the cost of switching vendors. Concentration among a handful of very large customers is the counterweight: it amplifies growth when those customers spend, and magnifies the impact when any one of them shifts its buying.

Memory: Feeding the Accelerators

Accelerators are only as useful as the data pipeline that supplies them. High-bandwidth memory (HBM) — stacked memory placed close to the processor — has become a critical component of leading-edge AI chips, and its supply has been a recurring constraint.

Micron Technology (NASDAQ: MU) is one of a small number of manufacturers capable of producing this memory at scale. Memory has historically been a brutally cyclical, commodity-like business, but the HBM segment behaves differently: it is technically demanding, closely tied to specific accelerator designs, and produced by only a few players. That scarcity has, at times, given HBM suppliers unusual leverage relative to the boom-and-bust reputation of the broader memory industry.

The caveat is that Micron is still a memory company. Its overall results are tied to the pricing cycle for DRAM and NAND, which can swing hard in both directions regardless of how tight the AI-specific segment is.

Why Bottleneck Suppliers Keep Their Leverage

The common thread across all four is that they sell into constraints. When demand outruns the supply of a critical, hard-to-substitute input, whoever controls that input sets terms. Long lead times, engineering integration, software lock-in, and limited numbers of qualified competitors all reinforce that position.

But the same characteristics that make these businesses powerful in an upcycle make them vulnerable in a downcycle. They are capital-intensive, they build capacity against demand forecasts that can prove too optimistic, and they are exposed to the spending decisions of a concentrated set of hyperscale customers. A slowdown in AI capital expenditure would hit the picks-and-shovels layer just as it hit the chipmakers.

What to Watch

For investors weighing these names, the questions worth tracking are less about any single quarter and more about the durability of the bottleneck. Does the power and cooling constraint ease as new capacity and grid connections come online? Does Ethernet hold its ground in AI networking, or do rival approaches gain? Does HBM stay scarce, or does added capacity erode its pricing premium?

The reward case rests on these constraints persisting long enough to justify today's expectations. The risk case is that the bottlenecks loosen, or that AI capital spending cools, and the leverage that lifted these suppliers works in reverse.