The Hidden Infrastructure Companies Behind the AI Boom

The AI boom is usually described through models, chips, and cloud platforms.

That description misses much of the physical system.

A high-end accelerator cannot operate without electricity arriving at the correct voltage. Thousands of accelerators cannot work together without high-speed connections. The heat they produce must be removed continuously, and the entire facility must keep running when equipment or utility power fails.

Behind every usable unit of AI computing capacity is a chain of transformers, switchgear, power electronics, cooling equipment, optical components, controls, and backup systems.

These products are less visible than an AI model, but they can determine when an expensive data center begins earning revenue.

That changes the AI business story. The opportunity is not limited to companies designing processors or selling software. It also reaches industrial manufacturers that can deliver reliable physical infrastructure at the scale and speed the buildout demands.

The AI boom has become an industrial buildout

McKinsey estimates that global data-center spending could reach approximately $7 trillion by 2030. Its definition of the equipment opportunity includes the electrical, thermal, and mechanical systems needed to build and operate a facility, excluding the semiconductors themselves (see Sources, [1]).

This distinction matters.

The chip is only one layer of the finished system. A functioning AI campus also needs utility connections, substations, transformers, switchgear, uninterruptible power systems, busways, backup generation, cooling loops, heat exchangers, network equipment, monitoring, and skilled installation.

The earlier Engines & Margins article The Hidden Economy Behind AI Data Centers mapped this wider system. The important business question is what happens when several of its least glamorous components become difficult to obtain.

A delayed chip order is visible immediately. A delayed transformer can be just as consequential because the completed servers may have no usable power.

That is why industrial equipment is moving from the background of the AI economy toward its critical path.

Power equipment can control the project schedule

Electricity does not travel directly from a transmission line into a GPU.

It passes through a sequence of equipment that changes voltage, distributes power, protects the facility, converts electricity, and keeps the site operating through faults or interruptions.

A shortage anywhere in this chain can delay the entire facility.

McKinsey reports that North American lead times can reach approximately 80 weeks for some medium-voltage switchgear and 50 weeks for large transformers (see Sources, [1]).

These delays create an unusual economic situation. A developer may already control the land, financing, building, and servers, yet still be unable to produce billable computing capacity.

The effect connects directly to the grid constraints examined in Why AI Data Centers Have a Power Problem Nvidia Can't Solve. Securing electricity is only one stage. The project must also obtain and install the equipment required to control and distribute it.

This makes delivery capacity valuable.

The strongest supplier is not necessarily the company with the most advanced individual component. It may be the company that can manufacture at scale, qualify its equipment for demanding customers, integrate it with adjacent systems, and provide a dependable delivery schedule.

Cooling is becoming its own equipment ecosystem

Nearly all the electricity consumed by computing equipment eventually becomes heat.

As more computing power is concentrated into each rack, removing that heat becomes more difficult. Traditional room-level air cooling remains useful, but high-density AI systems increasingly require cooling much closer to the processors.

Liquid cooling changes the equipment stack.

Instead of relying only on fans and chilled room air, a liquid system can include cold plates near chips, pumps, coolant-distribution units, manifolds, heat exchangers, sensors, connectors, leak detection, and facility-side heat-rejection equipment.

Each component creates another manufacturing, integration, monitoring, and maintenance requirement.

Bank of America forecasts cited by Reuters suggest that liquid cooling could rise from roughly 30% of new AI data-center installations today to approximately 70% by 2030 (see Sources, [2]).

That remains a forecast rather than a guaranteed outcome. Adoption will depend on rack density, cost, reliability, water and facility design, technical standards, and the pace at which operators replace or retrofit existing systems.

The broader direction is nevertheless important: cooling is becoming more tightly connected to the computing architecture.

A supplier can no longer design only for yesterday's server racks. It must understand how future chips will change power density, coolant temperatures, flow requirements, service procedures, and physical rack layouts.

The equipment around the processor must now follow the processor's road map.

Connectivity becomes part of computing performance

An AI cluster is not one enormous processor.

It can contain thousands of accelerators that repeatedly exchange data while training or serving a model. If those transfers are too slow, expensive computing hardware spends more time waiting and less time performing useful work.

The network therefore becomes part of the effective performance of the AI system.

This raises the importance of optical transceivers, lasers, photonics, controllers, switches, fiber, connectors, and the power components supporting those links.

STMicroelectronics provides one example of how this demand can affect a company outside the most visible group of AI chip designers. The company says its technologies support cloud AI through optical connectivity, power conversion, and control. It expected data-center revenue above $1 billion in 2026 and, subject to continued demand and customer engagements, above $2 billion in 2027 (see Sources, [4]).

Those figures are company expectations, not guaranteed results. Their importance is structural.

They show how AI infrastructure demand can reach semiconductor and industrial products that do not perform the model's main calculations. Moving data between processors can become a substantial business alongside manufacturing the processors themselves.

The AI hardware opportunity is therefore wider than the accelerator market. It includes the equipment that keeps accelerators supplied with power, within safe temperatures, and connected to one another.

Suppliers are moving from components toward systems

Customers building large AI facilities do not want hundreds of incompatible products that must be redesigned on site.

They increasingly need tested systems in which power, cooling, controls, and monitoring work together. This favors suppliers capable of integrating several layers or coordinating closely with other manufacturers.

Vertiv's announced agreement to acquire UtilityInnovation Group illustrates this movement.

Vertiv said the transaction involved approximately $1.45 billion in cash at closing, with up to another $1.15 billion tied to future performance targets. The company said the acquisition would add microgrid controls, onsite-generation orchestration, specialized switchgear, and behind-the-meter power architecture to its existing power-and-cooling portfolio (see Sources, [3]).

The strategic direction is more important than the transaction price.

A supplier historically focused on equipment inside the facility is moving further upstream toward the grid connection and onsite power. The intended result is a more coordinated chain from the source of electricity to the computing rack.

This reflects a wider change in customer priorities.

AI developers care about time to usable computing capacity. If one supplier or partnership can reduce design work, compatibility problems, installation delays, and commissioning risk, it may create value beyond the price of any individual component.

The product is no longer only a transformer, cooling unit, or power shelf. It is also the ability to make the complete system operational sooner.

Chip road maps now influence industrial road maps

Traditional industrial infrastructure is often designed around long planning and product cycles.

AI computing equipment can change much faster.

A new accelerator generation may increase rack power, alter cooling requirements, change electrical interfaces, or require different network bandwidth. Equipment suppliers must then redesign, test, qualify, manufacture, and service products compatible with that new architecture.

McKinsey argues that chip and reference-architecture road maps increasingly define downstream power, thermal, and interconnection requirements. Suppliers may need to work with semiconductor companies and hyperscalers earlier instead of waiting for a conventional equipment request (see Sources, [1]).

This compresses two different industrial clocks.

Semiconductor platforms can change within months. Transformers, switchgear, cooling systems, factories, certification processes, and field-service networks have traditionally evolved over years.

The companies that handle this mismatch may gain an advantage. They need enough manufacturing discipline to deliver mission-critical equipment reliably, but enough engineering speed to keep pace with changes in computing.

Scale without adaptability risks producing yesterday's equipment.

Adaptability without scale risks producing an impressive prototype that cannot support a gigawatt-scale rollout.

Higher-voltage DC could rearrange the supplier landscape

Today's power chain involves several conversions before electricity reaches the components inside a server.

At high rack densities, each conversion can add losses, heat, equipment, and physical complexity. One proposed response is to distribute power through parts of the facility at approximately 800 volts direct current.

McKinsey says AI-focused rack requirements that were once below 50 kilowatts are rising toward 200 to 600 kilowatts and could approach one megawatt later in the decade. It argues that an 800 VDC architecture could move similar power at roughly half the current of common facility-level AC approaches while reducing some conversion stages (see Sources, [5]).

If adoption grows, demand could shift.

Some conventional low-voltage equipment may be redesigned or reduced, while opportunities expand for DC-rated busways, protection systems, breakers, connectors, power sidecars, power semiconductors, and potentially solid-state transformers.

This is not a completed transition.

Higher-voltage DC introduces new safety, fault-protection, training, maintenance, standardization, and interoperability requirements. Existing facilities may be expensive to retrofit, and operators need confidence that new mission-critical equipment can remain reliable for decades.

The direction nevertheless demonstrates why current demand does not guarantee permanent leadership.

The architecture can change the equipment mix. A company benefiting from today's bottleneck must continue investing so that its products remain relevant to the next system design.

How infrastructure suppliers make money

AI infrastructure suppliers can generate revenue through several related activities:

• Selling physical equipment
• Designing equipment for a customer's architecture
• Integrating power, cooling, and control systems
• Commissioning equipment at the site
• Providing replacement parts and maintenance
• Monitoring installed systems
• Expanding an existing facility
• Supporting upgrades to higher-density computing

The mix differs by company.

A component manufacturer may rely heavily on factory volume. A systems supplier may earn more from engineering, integration, software, and services. A company with a large installed base may benefit from maintenance and replacement demand long after the original construction project ends.

This creates several possible competitive advantages.

Manufacturing capacity matters when customers need thousands of units. Qualification matters because operators cannot accept frequent failures. Integration matters because every subsystem must work with the others. Service coverage matters because downtime in an AI cluster can be extremely expensive.

The business opportunity is therefore not simply producing more hardware. It is becoming difficult to replace inside a customer's infrastructure plan.

Why every supplier will not automatically win

A rapidly growing market can still create poor business outcomes.

Suppliers may expand factories just as demand slows. Competitors can add capacity and reduce pricing power. A new architecture can weaken demand for an older product. Large customers may demand lower prices or redesign systems around alternative vendors.

Project timing creates another risk.

Orders can be delayed, changed, or cancelled when grid access, permits, financing, or customer demand shifts. A large backlog does not always convert into revenue on the original schedule.

Technical execution also matters. Cooling leaks, electrical failures, delayed qualification, poor integration, or insufficient field support can damage customer relationships in a market where reliability is essential.

This is why the hidden-supplier theme should not be reduced to a list of stocks.

The useful question is which capabilities are becoming scarce and which companies can supply them reliably while the architecture continues changing.

What to watch across the supplier ecosystem

Several signals can reveal where value is moving:

1. Lead times: Persistent delays indicate that manufacturing capacity remains constrained.
2. Qualification: Winning approval for a hyperscaler's reference architecture can matter more than announcing a prototype.
3. Manufacturing expansion: New factories indicate confidence but also raise the risk of future overcapacity.
4. Integration: Partnerships and acquisitions show suppliers attempting to control more of the complete system.
5. Installed-base services: Maintenance and upgrade revenue can make the business less dependent on new construction.
6. Architecture exposure: Suppliers must show how their products fit liquid cooling, optical networking, microgrids, and higher-voltage DC.
7. Customer concentration: Fast growth tied to only a few buyers can make revenue more fragile.
8. Margins and cash flow: Rising orders create value only when the supplier can deliver profitably.

These indicators separate infrastructure demand from durable business quality.

The AI Data Center Power & Cost Calculator can help illustrate why increases in computing load, utilization, overhead, and PUE create such large facility-level power and cooling requirements.

The companion article How Data Centers Actually Make Money explains how operators turn that constructed capacity into recurring revenue.

The physical businesses underneath artificial intelligence

AI may appear weightless when it is accessed through a browser.

Its infrastructure is not.

Every new cluster requires electricity to be transformed, controlled, converted, and distributed. Its heat must be captured and removed. Its processors must exchange data through increasingly sophisticated networks. Its site must keep operating when individual systems fail.

That creates businesses far beyond the companies designing models and accelerators.

Transformers, switchgear, cooling systems, optical links, controls, power electronics, and microgrids may look like supporting equipment. In practice, they can determine when a facility opens, how efficiently it runs, and how much computing capacity becomes usable.

The AI boom is a software story built on an industrial foundation.

Some of its most important companies may be the ones manufacturing that foundation.