Are we building tomorrow’s infrastructure or yesterday’s data centers?

Key Takeaways

Ask an artificial intelligence tool a question and the answer appears in seconds. Behind that simple exchange, processors, networks, power systems and cooling equipment all work together. That’s why a technology that we experience as software is producing such a visible construction boom. But the buildings are only part of the story. As AI changes the work computers do, it also changes the machinery inside them. And the machinery changes are happening much faster than the facilities built to house them.

We’ve seen this before, but never this fast

Around the turn of the 20th century, many factories relied on a central steam engine or waterwheel, with shafts and belts carrying power to individual machines. That arrangement determined where equipment could go.

Electric motors loosened those constraints. Machines could have their own source of power and be arranged around the flow of production. Old and new systems existed alongside one another as businesses adapted, but the factory floor itself changed and the buildings had to change. The way people worked had to change.

Electrification took decades, but this is happening in years. The machinery inside data centers is evolving faster than the buildings can be built. Technology that’s standard today is obsolete in five years.

From storing information to generating it

Data centers already supported online banking, streaming, cloud applications, healthcare and government services before the current AI boom. AI adds something different: generating new text, images, audio, video and other outputs through repeated calculations.

Two terms explain the demand. Training is developing a model by adjusting it using data. Inference is what happens when the trained model handles a new request. Training requires large groups of processors working together. Inference creates ongoing computing demand every time people use the model.

A single question is only part of the picture. Demand grows when AI is built into search, office software, customer service and other tools used throughout the day. Systems that generate video or carry out multiple tasks require far more computation than a short text response.

The Department of Energy estimated that data centers could consume 6.7% - 12% of U.S. electricity by 2028, up from 4.4% in 2023. The International Energy Agency projects that global data center electricity consumption will roughly double between 2025 and 2030. 

The machinery is changing rapidly

AI changes how computers work together. Graphics processing units perform many calculations at once, making them well suited to modern AI work. In large systems, groups of these processors need fast connections to share information and coordinate their work.

More computing capacity packed into a rack means more electricity and more concentrated heat, and cooling becomes an issue. Fluid flowing close to the computing components carries away the heat in a process called direct liquid cooling. Other designs immerse equipment in specialized liquids. These systems improve heat removal significantly, but their water consumption depends on the full cooling system.

The cooling technology that’s standard today is not the cooling technology that will be standard in five years. Facilities built to current specs won’t perform as well. Data centers built 10 years ago are being retooled or shut down because the technology has moved on. That’s not a prediction, it’s what’s happening now.

Computing chips get more powerful and use less electricity every year. A data center that’s cutting edge today is middle of the pack in three years, obsolete in seven. That’s the lifecycle of data center technology.

Efficiency doesn’t stop the building boom

It seems reasonable to expect that better chips and software would reduce the need for new facilities. They reduce the resources needed for a particular task, but that’s only half the equation.

When computing becomes more affordable and useful, people find more ways to use it. Some applications demand more resources. A system that checks several sources and completes multiple steps does more work than one that returns a brief answer.

The International Energy Agency’s analysis describes rapid improvements in energy efficiency per AI task. Even so, it projects that global data center electricity consumption will roughly double by 2030. Efficiency is improving while the volume and complexity of the work are growing.

The next chapters will have more than one shape

Not every AI task requires a giant centralized facility. Some processing happens on phones, laptops, or smaller systems closer to where people are. Large facilities remain important for work that demands substantial computing capacity. These approaches coexist, with different tasks handled in different places.

The industry is moving toward distributed computing. Processing happens closer to where data originates instead of sending everything to one massive facility far away. It’s faster, cheaper, and uses less electricity. That shift is already underway.

Quantum computing is being developed to complement conventional computing for certain problems. It doesn’t replace the computers delivering everyday digital services.

The computing infrastructure of tomorrow won’t look like the computing infrastructure of today. The massive, centralized facility built to current specifications may remain valuable. Or it becomes substantially different. Or it’s in the wrong place. The technology itself is evolving faster than anyone can predict with certainty.

What counties are actually seeing

This technology shifts shows up for counties in the boom in data center proposals. The demand is real and driven by genuine computing needs. Every digital service people rely on, from healthcare to government to e-commerce to AI tools, depends on data center infrastructure somewhere.

If a county declines to host that infrastructure, the computing still happens elsewhere. The jobs, the tax revenue, the economic activity and the infrastructure investment move to communities that said yes. For counties that want to participate, construction brings immediate work, operations employ specialized technicians and engineers and the facility signals to other tech-dependent businesses that the area can support their infrastructure needs.

The facility being proposed today will not be the facility operating in a decade. The cooling technology will change. The electricity demand will shift. The computing model will be different. Efficiency improvements already in development will reshape what’s actually needed.

Committing electricity, water, and land use to a facility based on current technology means committing to something that’s already becoming obsolete. That’s not an argument against data centers. It’s a reality about what you’re committing to.

Three things counties should understand

The electricity-to-factories parallel is instructive. That shift took decades, but this one is happening in years. Infrastructure commitments counties make today are built for a computing world that’s already changing.

• Infrastructure mismatch - Grid and water infrastructure built today won’t match what the facility actually needs in five years. Cooling technology changes. Efficiency improves faster than projected. Electricity demand shifts. What was built to spec becomes oversized, undersized, or simply wrong.
• Lifecycle mismatch - Facility lifecycles are 10 to 15 years, not 30 years. Data centers lose value faster than infrastructure costs are recovered. When the facility loses value to the developer, counties hold infrastructure that doesn’t serve its purpose.
• Location mismatch - Computing is moving toward distributed models, not massive centralized facilities. Siting decisions that make sense today won’t make sense in a decade. Current location assumptions become the wrong location.

Counties can still welcome data center development, but they need to do it from a position of understanding what the landscape is becoming. The machinery, the efficiency, the location, the cooling technology, the entire computing model will be different. That’s not speculation, that’s the current trajectory.