AI Data Center Infrastructure: What Business Leaders Need to Know About Power, Cooling, Networks and Scale

A fact-based guide for business leaders, policymakers, utilities, developers and communities evaluating the infrastructure behind artificial intelligence.

Artificial intelligence may feel like a digital technology, but the infrastructure supporting it is becoming increasingly physical.

Behind every large AI model are processors, memory, storage systems, high-speed networks, electrical substations, transformers, cooling plants, fiber routes and, increasingly, dedicated sources of power. As AI adoption expands, these systems are becoming some of the largest infrastructure projects being built anywhere in the world.

That scale is changing the public conversation around data centers.

Communities are asking whether AI data centers will strain electric grids, consume scarce water, generate excessive noise or increase residential utility bills. Utilities are trying to determine how quickly they can connect extraordinarily large new loads. Policymakers are weighing economic development against infrastructure costs and environmental impacts. Enterprises are confronting another question: whether enough computing infrastructure can be built quickly enough to support the AI services on which their future strategies increasingly depend.

These concerns are legitimate. But they are also frequently oversimplified.

There is no single standard design for an AI data center. Two facilities with similar computing capacity can have very different impacts depending on their cooling architecture, source of electricity, geographic location, water strategy, grid connection, network infrastructure and proximity to residential areas.

Understanding those differences is becoming important well beyond the technology industry.

AI Is Turning the Data Center Into Critical Industrial Infrastructure

The scale of the change is becoming difficult to ignore.

Lawrence Berkeley National Laboratory's latest U.S. data-center analysis estimates that data centers could account for approximately 11.8% of total U.S. electricity consumption by 2030 in its reference case. Because the trajectory of AI hardware, utilization and deployment remains uncertain, the laboratory estimates a broader range of approximately 9.5% to 15.3%. Its reference case equates to roughly 649 terawatt-hours of electricity annually. Lawrence Berkeley National Laboratory — United States Data Center Energy Usage Report: 2025 Update

The trend is global as well. The International Energy Agency estimates that data centers consumed approximately 415 TWh of electricity worldwide in 2024, or around 1.5% of global electricity consumption. In its base case, the IEA projects that consumption could more than double to approximately 945 TWh by 2030. Electricity consumption by accelerated servers—the systems most closely associated with AI—is projected to grow particularly rapidly. International Energy Agency — Energy and AI

Those numbers include more than AI alone, and forecasts this far into a rapidly changing technology cycle should not be treated as precise predictions. But the direction is clear: computing is becoming a major new source of electricity demand.

That makes the modern AI data center less like an ordinary commercial building and more like a highly specialized industrial facility.

The Data Center Building Is Not Really the Starting Point

Traditional commercial construction generally begins with a building and then determines what electrical and mechanical systems are required to support it.

AI infrastructure increasingly works in the opposite direction.

The design begins with the computing workload.

How many accelerators will operate in the facility? How much electricity will they require? How densely will they be installed? What network latency is acceptable? What availability level is required? How much heat will the equipment generate?

The answers determine almost everything that follows.

Modern AI systems illustrate just how dramatically compute density has changed. NVIDIA's DGX GB200 NVL72 rack-scale architecture integrates 72 Blackwell GPUs and 36 Grace CPUs. NVIDIA documentation places rack power consumption at approximately 120 kilowatts, with direct liquid cooling built into the architecture. NVIDIA DGX GB200 User Guide

A single rack drawing that much electricity presents an entirely different engineering problem from the racks that populated conventional enterprise server rooms.

And virtually all of the electricity consumed by those processors, memory systems and networking components ultimately becomes heat.

A 100-megawatt IT environment therefore creates not only a 100-megawatt-scale electrical requirement but also a correspondingly enormous thermal-management problem.

That is one reason cooling architecture has become central to AI data-center design.

Why Liquid Cooling Is Becoming So Important

For decades, data centers were associated primarily with air cooling: cold air was delivered to servers, warm air was removed, and chillers or other systems ultimately rejected that heat outside the facility.

Air cooling is not disappearing. But as rack density rises, moving enough heat with air alone becomes increasingly difficult.

Liquid has far greater heat-transfer capability than air. Direct-to-chip systems therefore bring coolant much closer to the processors generating the heat. Cold plates attached to CPUs, GPUs and other components absorb thermal energy, while pumps move heated fluid back through the cooling system.

The important distinction is what happens next.

A liquid-cooled data center does not necessarily consume large quantities of water simply because water or another liquid is circulating through the equipment.

In a closed-loop system, the coolant can circulate repeatedly.

That distinction between water circulating inside a system and water actually being consumed is critical to understanding modern data-center water use.

Data Centers Can Consume Significant Water—but the Cooling Design Matters Enormously

Traditional evaporative cooling can be remarkably effective.

Cooling towers remove heat partly by evaporating water into the atmosphere. From an energy-efficiency perspective, that can be attractive. But the evaporated water is consumed and must be replaced.

That creates an obvious concern in regions where freshwater is scarce.

Lawrence Berkeley National Laboratory estimated direct U.S. data-center water consumption at approximately 66 billion liters in 2023, with hyperscale and large colocation facilities accounting for most of the total. But that national number masks enormous variation among facilities because cooling technology, local climate and operating practices differ substantially. LBNL — 2024 United States Data Center Energy Usage Report

New cooling architectures demonstrate just how large that variation can become.

Microsoft introduced a data-center design optimized for AI workloads that uses direct-to-chip cooling in a closed loop and requires zero water evaporation for cooling during normal operations. Once the system is filled, water circulates between servers and the cooling infrastructure rather than continually being evaporated and replenished. Microsoft estimates that the design can avoid more than 125 million liters of water use annually per data center compared with the company's FY2024 global-average water-use baseline. Microsoft — Next-generation datacenters consume zero water for cooling

This does not make the entire facility "water free." Water can still be required for sanitation, commissioning, maintenance and other purposes. Electricity generation can also have an indirect water footprint depending on the generation source.

That last point matters. The U.S. Geological Survey notes that many thermoelectric generating plants themselves withdraw and consume water for cooling. A complete water assessment should therefore consider both direct water use at the data center and indirect water use associated with its electricity supply. U.S. Geological Survey — Thermoelectric Power Water Use

This is why blanket statements such as "AI data centers use enormous amounts of water" are not particularly useful without knowing the facility design.

Some do.

Others can be designed to use remarkably little water for operational cooling.

Conserving Water Can Create an Energy Tradeoff

There is another complication: minimizing water consumption does not automatically minimize electricity consumption.

Evaporative cooling can reject heat with relatively low electrical energy. Replacing it with dry or mechanical cooling may increase electricity requirements, particularly during hot weather.

Microsoft itself acknowledges this tradeoff in its zero-water cooling architecture, noting that eliminating evaporation can increase Power Usage Effectiveness, although higher-temperature liquid cooling and more efficient chillers can offset part of that penalty. Microsoft — zero-water cooling architecture

This is an important engineering principle.

There is rarely one metric that defines a "green" data center.

Power Usage Effectiveness, or PUE, measures total facility energy divided by the energy consumed by IT equipment. A facility using 120 MW to support 100 MW of IT equipment would have a PUE of 1.20. The U.S. Department of Energy uses PUE and related measures such as Water Usage Effectiveness as important data-center efficiency metrics. U.S. Department of Energy — Best Practice Guide for Data Center Design

But improving one metric can sometimes make another worse.

Water, electricity, carbon, reliability and cost have to be optimized as a system.

Geography therefore matters enormously.

Power May Become the Defining Constraint on AI Growth

AI processors can be manufactured and deployed much faster than new transmission lines, substations and generating plants can typically be permitted and constructed.

That mismatch is becoming one of the defining infrastructure challenges of the AI era.

A large AI campus may require hundreds of megawatts of continuous power. At that scale, the electrical infrastructure can include high-voltage transmission, utility-scale substations, transformers, medium-voltage campus distribution, switchgear, batteries, uninterruptible power systems and multiple levels of redundancy.

Increasingly, developers are also considering power generation on the campus itself.

EPRI's DCFlex program reported in September 2026 that grid-interconnection delays and transmission constraints are making on-site generation increasingly important. EPRI specifically identifies backup power, peaking generation, bridge power and even primary on-site generation as emerging elements of data-center planning. EPRI DCFlex research on data-center grid integration

Bridge power is particularly important.

Imagine a data center that can be completed in two years but cannot receive its full utility allocation for another four or five years. Temporary or transitional on-site generation can allow the computing infrastructure to begin operating while permanent grid infrastructure catches up.

That does not necessarily mean the temporary configuration becomes the facility's long-term energy architecture.

It may simply be a way of reconciling two industries operating on very different construction timelines.

Microsoft's 2-Gigawatt Pecos Project Shows Where the Market Is Heading

Microsoft's recently announced AI data-center campus in Pecos, Texas provides an unusually clear illustration.

Microsoft says the development will add approximately 2 gigawatts of data-center capacity and represents a multibillion-dollar investment expected to unfold over five to seven years. Microsoft — Pecos AI datacenter announcement

What makes the project particularly interesting is the infrastructure surrounding the computing capacity.

At launch, Microsoft plans to operate the campus with a co-located natural-gas power facility behind the meter. Microsoft says it is funding both the new generation and the supporting energy infrastructure, and that the campus will initially operate independently from the public grid. The company expects the power facility and data center eventually to connect with the broader regional grid.

Its cooling strategy is equally notable. Microsoft says the Pecos facility will use a closed-loop cooling system requiring an initial charge of fluid but no additional cooling-water consumption during steady-state operation, with nonpotable water used where possible. Microsoft — Pecos AI datacenter announcement

The official announcement does not disclose the precise initial-fill volume, so it would be inappropriate to present a speculative figure as fact.

Pecos also illustrates an important distinction for public policy. Behind-the-meter generation can reduce the immediate requirement placed on an existing electric grid, but it does not eliminate environmental or infrastructure considerations. Natural-gas generation still requires fuel, air permitting, emissions controls and supply infrastructure. Microsoft says the Pecos generation plant will employ selective catalytic reduction to reduce nitrogen-oxide emissions.

The correct question is therefore not simply whether the facility "uses the grid."

The question is how the entire energy system has been engineered and who bears its costs and impacts.

Will AI Data Centers Raise Everybody's Electric Bill?

This is one of the most important questions communities are asking, and the correct answer is neither an automatic yes nor an automatic no.

A very large new electrical customer can require additional generation, transmission and substations. Those assets cost money. If infrastructure is built for one large customer and the associated financial risk is shifted to the wider rate base, other customers can be affected.

The U.S. Department of Energy has therefore identified fair cost allocation, stranded-asset risk and system reliability as central issues in designing electric rates for data centers and other exceptionally large loads. U.S. Department of Energy — Electricity Rate Designs for Large Loads

Virginia provides a useful case study in how nuanced this can become.

The state's Joint Legislative Audit and Review Commission found that existing data-center customers were paying their allocated cost of service under current utility structures. At the same time, JLARC warned that the enormous amount of new generation and transmission required to accommodate continued growth could increase electricity-system costs more broadly. It recommended considering mechanisms such as distinct customer classifications and cost-allocation changes to better insulate other customers. Virginia JLARC — Data Centers in Virginia

Federal regulators are now addressing the issue as well. In June 2026, the Federal Energy Regulatory Commission directed the six regional transmission organizations under its jurisdiction to justify or reform aspects of their tariffs governing data centers and other large loads. U.S. Department of Energy — FERC action on large-load interconnection reform

For policymakers, this is a much more productive framework than debating whether data centers inherently increase residential bills.

The better questions are: Who is paying for the new infrastructure? What happens if projected demand fails to materialize? Who bears stranded-asset risk? Is the customer committing to minimum demand? Is dedicated generation being added? Can the facility provide grid flexibility?

The answers are contained in rate structures, interconnection agreements and infrastructure commitments—not in the headline megawatt number alone.

The AI Data Center Is Starting to Resemble a Power System

Combine utility service, on-site generation, batteries, backup systems, cooling infrastructure and sophisticated control software, and a large AI campus can begin to resemble a microgrid.

That architecture also creates the possibility of greater flexibility.

EPRI's DCFlex initiative is studying whether data centers can become more responsive participants in the electric system rather than behaving as completely inflexible loads. Flexibility does not necessarily mean shutting down a major AI training run whenever the utility needs power. It can come from batteries, thermal systems, generators, scheduling less urgent workloads or shifting some compute geographically. EPRI DCFlex — Workstream 2

This could eventually turn some data centers from purely large electrical consumers into sophisticated energy-system participants.

But that concept is still evolving, and the amount of flexibility available will depend heavily on workload, architecture and contractual arrangements.

AI Is Also Reviving Interest in Firm Power, Including Nuclear Energy

Wind and solar can provide extremely valuable low-carbon electricity, and large technology companies have financed substantial renewable development through power-purchase agreements.

But AI data centers operate continuously.

That makes firm power—electricity that is predictably available when required—especially valuable.

This helps explain growing technology-industry interest in nuclear generation.

Constellation and Microsoft entered a 20-year power-purchase agreement supporting the planned restart of the former Three Mile Island Unit 1, renamed the Crane Clean Energy Center, with approximately 835 MW of output. Constellation — Crane Clean Energy Center announcement

Google, meanwhile, has entered agreements involving Kairos Power's advanced nuclear technology. The first planned project with the Tennessee Valley Authority would supply 50 MW beginning around 2030, with Google's broader Kairos relationship targeting as much as 500 MW across multiple deployments. Google — advanced nuclear project with Kairos Power and TVA

Advanced nuclear is not a near-term solution for every data-center project. Permitting, financing, construction timelines, fuel supply and technology maturity remain significant considerations.

But the underlying economic signal is important: AI is creating demand for large quantities of highly reliable electricity on a scale sufficient to influence investment in generation technologies themselves.

Noise Is a Real Issue—and Largely a Siting and Engineering Issue

Not every data-center impact occurs at the regional or national level.

For nearby residents, noise can be much more immediate.

Cooling fans, transformers, pumps, chillers and generator testing can all create sound. Low-frequency or tonal sound can also be more noticeable than an equivalent amount of broadband background noise.

Virginia's JLARC review concluded that noise has created problems at some sites located near residential areas, while also observing that the large majority of data centers do not generate noise complaints because of their location or design. The report recommended improved sound modeling and zoning approaches that account for continuous and low-frequency noise. Virginia JLARC — Data Centers in Virginia

That distinction matters.

Noise is not an unavoidable property of computing. Site orientation, mechanical-equipment placement, setbacks, acoustic barriers, enclosures, berms, fan selection and operating schedules can all materially affect what surrounding communities experience.

The best time to address these issues is during site selection and design—not after construction.

Site Selection May Be the Most Important Decision of All

Many of the disputes surrounding data centers are ultimately siting problems.

An enormous computing campus placed close to residential neighborhoods creates a fundamentally different planning challenge than one developed in a large industrial zone with substantial setbacks.

The ideal site must satisfy several constraints simultaneously.

It needs access to enormous amounts of electricity or a credible plan to develop them. It needs multiple high-capacity fiber routes, preferably with true physical diversity. It requires appropriate land, zoning and permitting. Cooling must fit the climate and local water environment. Electrical and mechanical equipment must be positioned so that noise can be managed. The surrounding transportation network must support years of construction. Future expansion must also be considered.

This is why inexpensive land alone does not make a good AI data-center site.

Power, network connectivity and infrastructure readiness increasingly determine the value of the land.

This also reinforces a broader theme we have discussed at Macronomics: AI may be digital, but its limiting factors are increasingly physical. As we explored in our analysis of the AI Memory Wall, adding more compute eventually exposes constraints in memory, networking, power, cooling and physical infrastructure. Likewise, our work on AI moving the bottleneck highlights how advances in one part of the AI ecosystem can simply shift the constraint somewhere else.

AI infrastructure is increasingly a systems-engineering problem.

What Should Communities Actually Ask Before Approving an AI Data Center?

The most useful public discussion is not whether data centers are universally "good" or "bad." It is whether a specific project has been designed and financed appropriately for a specific location.

Local governments, regulators and economic-development officials should understand at least six issues before evaluating a major project:

·         Power: Where will the electricity come from, what infrastructure must be constructed, and who will pay for it?

·         Water: What cooling architecture will be used, how much water will actually be consumed, what type of water will be used, and how does that compare with local availability?

·         Noise and land use: What are the modeled nighttime sound levels, equipment locations and setbacks from sensitive properties?

·         Network infrastructure: Are multiple physically diverse fiber routes available, and will public infrastructure require expansion?

·         Economics: What tax revenue will remain after incentives, and what infrastructure obligations is the developer assuming?

·         Risk: What protections exist if the project is delayed, downsized or abandoned after utilities have made major capital investments?

Those questions are far more useful than evaluating a proposal solely by acreage or megawatts.

The Economic Benefits Can Be Significant—but They Need to Be Structured Correctly

Data centers are extraordinarily capital-intensive, but they are not traditional factories.

Virginia's JLARC found that a typical 250,000-square-foot data center may support only around 50 full-time operational workers, although construction can involve approximately 1,500 workers at peak activity. Much of the economic impact therefore occurs during construction and through capital investment rather than through thousands of permanent employees. Virginia JLARC — Data Centers in Virginia

Tax revenue can nevertheless be substantial.

Among five Virginia localities with mature data-center markets, JLARC found that data-center revenue ranged from less than 1% to as much as 31% of total local revenue, depending on the locality and its tax structure. Virginia JLARC — Data Centers in Virginia

That makes incentive policy extremely important.

A community that gives away too much of the tax base while absorbing new infrastructure costs can turn an apparently attractive development into a poor economic bargain. A community with scarce, highly suitable data-center land, existing power infrastructure and strong fiber connectivity may have considerably more negotiating leverage.

The objective should be shared value: the developer receives a viable site and predictable approval process, while the community protects residents, avoids inappropriate infrastructure subsidies and retains an appropriate portion of the economic benefit.

The Bigger Lesson: There Is No Such Thing as “The AI Data Center”

The public debate often treats AI data centers as though they were interchangeable.

They are not.

A facility dependent on continuous evaporative cooling in a water-stressed region is different from one using closed-loop liquid cooling.

A 500-megawatt campus requiring extensive utility expansion is different from one accompanied by dedicated generation that the developer finances.

A data center built close to homes is different from one appropriately buffered inside an industrial district.

A facility connected through one constrained fiber corridor is different from a campus with geographically diverse network infrastructure.

And a municipality that gives away most of the project's tax value is in a very different economic position from one that carefully structures infrastructure commitments, rate protections and long-term revenue.

As AI infrastructure expands, businesses and governments will increasingly need to evaluate these projects at the system level.

The critical questions are not simply how large the data center will be or how much power it will consume. They are how the facility will be powered, cooled, connected and financed; what risks it creates; who bears those risks; and what long-term value the project creates for enterprises, utilities and communities.

The AI infrastructure buildout is real, and its scale is extraordinary.

But its consequences are not predetermined.

Good engineering, appropriate siting, disciplined economic analysis and thoughtful public policy can make the difference between an infrastructure project that creates conflict and one that creates durable strategic value.

Macronomics Perspective

At Macronomics, we view AI infrastructure as a business-transformation issue as much as a technology issue. The rapid expansion of AI is connecting decisions that were once evaluated separately—including compute architecture, network design, cloud connectivity, carrier services, energy, facilities, capital planning, regulation and enterprise AI strategy.

For business leaders, investors, developers and public-sector decision makers, understanding those interdependencies is becoming essential. The organizations that succeed in the AI economy will not simply acquire more AI. They will build the network, infrastructure and operating models required to support it securely, reliably and at scale.

Macronomics helps organizations evaluate AI strategy, infrastructure readiness, and the network, cloud, and communications technologies required to support and scale artificial intelligence.