News

Power, Water, Jobs, Taxes: Weighing the AI Buildout's Real Costs and Benefits

• By John K. Waters
• 07/17/2026

New York Governor Kathy Hochul did something on July 14 that would have been difficult to imagine at the start of the generative AI boom. She told the data center industry to stop.

Hochul signed an executive order imposing a one-year moratorium on certain new hyperscale data centers, making New York the first state to pause the development of these facilities statewide. The order temporarily halts state environmental permitting for new data centers requiring more than 50 megawatts of electricity while regulators study their effects on the power grid, water supplies, ratepayers and host communities. Previously approved projects were not broadly canceled.

The announcement landed like a thunderclap in an industry accustomed to states competing for its business with tax exemptions, expedited permitting and promises of cheap land and power.

President Donald Trump attacked the decision, writing on Truth Social that Hochul had made a "terrible decision" that would send investment to other states, and that the taxes and jobs data centers bring "amount to LIQUID GOLD." Hochul answered that New York was not rejecting data centers outright. It was demanding better terms.

"But when you benefit from the talent and energy of New York, we expect you to protect our resources and give back to our communities," Hochul said at a press conference announcing the order. In a written statement, she put it this way: "New York will lead the way in creating the strongest standards in the nation for data center development, ensuring that when companies succeed because of New York, New Yorkers succeed too."

The exchange crystallized a political struggle spreading across the United States. Artificial intelligence may live in the cloud, but the cloud has become impossible to ignore on the ground.

It looks like acres of windowless buildings, electrical substations, transmission lines, cooling equipment, backup generators, and construction traffic. It draws electricity continuously, sometimes at a scale comparable to a city. Depending on how it's cooled, it can also require substantial water. It may produce billions of dollars in investment, but comparatively few permanent jobs.

The question confronting communities is no longer whether AI needs infrastructure. It does. The harder question is whether the people living next to that infrastructure are being asked to assume too much of its cost.

A conventional image of a data center is a warehouse filled with computer servers. That description is technically correct, but increasingly misleading.

The largest AI facilities now resemble industrial plants in their resource demands. A 100-megawatt data center running continuously consumes 876,000 megawatt-hours of electricity a year. A 1-gigawatt campus would use ten times that amount, roughly the output of a large nuclear reactor operating at full capacity.

Unlike a stadium, which experiences short bursts of demand during games and concerts, a data center is designed to operate around the clock. The stadium is a sprinter. The data center is a marathon runner that never stops.

The nationwide effect is beginning to show. A report from the U.S. Department of Energy and Lawrence Berkeley National Laboratory estimated that data centers consumed about 4.4 percent of U.S. electricity in 2023. Their share could rise to between 6.7 percent and 12 percent by 2028.

New York's own queue illustrates why officials are nervous. Hochul's executive order says nearly 12 gigawatts of proposed data center load was waiting to connect to the state's electric system as of May 2026, with more than 8 gigawatts entering the queue during 2025 alone. Twelve gigawatts is approximately the output of a dozen large nuclear reactors. Not every proposed project will be built, but utilities may still need to plan transmission and generation around requests that could materialize.

That creates an unusual planning problem. If utilities build substations, transmission lines or generation for projects that never arrive, ordinary customers could be left paying for stranded infrastructure. If utilities wait until demand is certain, the data centers may not be able to connect for years.
Hochul's order explicitly states that upgrades needed to serve large new loads should not be financed by "every-day New Yorkers." It directs regulators to develop rules to assign more of those costs to the facilities that create them.

Comparisons are imperfect, but they help explain the scale.

A large 100-megawatt data center could draw roughly as much continuous power as tens of thousands of homes. A 1-gigawatt campus can approach the electricity demand of hundreds of thousands of households, depending on regional residential consumption.

A major stadium can draw a tremendous amount of electricity during an event. Floodlights, video boards, kitchens, refrigeration, ventilation, broadcasting equipment and tens of thousands of spectators all contribute. Yet an NFL stadium hosts a limited number of games each year, along with concerts and other events. Its annual consumption is generally measured in millions or tens of millions of kilowatt-hours.

A 100-megawatt data center running continuously consumes 876 million kilowatt-hours a year.

The difference is not merely size. It's persistence.

A steel mill, oil refinery, semiconductor fabrication plant, or aluminum smelter is a closer power comparison because those facilities also operate continuously and can draw hundreds of megawatts. The difference is what they produce and how communities value that output.

A factory produces cars, steel, chips or chemicals. It may directly employ hundreds or thousands of people. A hyperscale data center provides computing capacity that can support cloud services, business software, scientific research, streaming media, government systems and AI. Its economic output may be enormous, but much of that value is created remotely, and the number of people working inside the facility after construction can be relatively small.

This helps explain the resentment. Communities are being asked to host infrastructure comparable in scale to heavy industry without always receiving a heavy industry employment base.

Water consumption has become one of the loudest arguments against data centers, but it's also one of the easiest to oversimplify.

Data centers can cool their equipment in several ways. Evaporative cooling can reduce electricity consumption, but it consumes water. Air-cooled or dry-cooled systems use less direct water, but may require more electricity, particularly during hot weather. Closed-loop liquid cooling can recirculate coolant, although heat must still ultimately be removed from the system.

As a result, two facilities with similar computing capacity can have dramatically different water footprints.

Researchers at the University of California, Riverside, examining U.S. water systems in a peer-reviewed analysis led by associate professor Shaolei Ren, estimated that, if 2024 water-use intensity continued, data center expansion through 2030 could require between 697 million and 1.45 billion gallons per day of new water-system capacity. At the high end, that is comparable to, or greater than, New York City's average daily water supply. The researchers emphasized that the effect would not be evenly distributed. It would be concentrated in the communities hosting the facilities, often during the hottest periods when water and electricity systems are already under stress.

The golf course comparison is particularly useful.

According to a 2025 irrigation survey conducted by the Golf Course Superintendents Association of America and funded in part by the United States Golf Association, irrigation demand varies enormously by region, rainfall, turf acreage and water source. A golf course in a humid climate may require far less irrigation than one in the desert Southwest. Nationally, most courses draw from wells, ponds or recycled water rather than municipal drinking-water systems, and total U.S. golf irrigation demand has fallen 31 percent since 2005.

A single large, water-cooled data center can consume water in the same broad range as one or several irrigated golf courses. A sprawling hyperscale campus could use substantially more. Yet the comparison changes by location. A desert golf course may be the heavier water user, while a data center built with dry cooling may consume relatively little direct water.

A stadium generally uses much less water than either a heavily irrigated golf course or a water-cooled hyperscale data center. It needs water for restrooms, concessions, cleaning, cooling and field maintenance, but its occupancy is intermittent. Stadiums have also adopted low-flow fixtures, recycled water and more efficient cooling. Allegiant Stadium, for example, says it meters cooling towers and other water systems and has saved an estimated 535 million gallons of water through efficient fixtures.

The broadest fair comparison is this: data centers are not uniquely thirsty by the standards of heavy industry, agriculture, semiconductor manufacturing or desert recreation. What alarms communities is the prospect of adding several of them at once, often without transparent information about peak demand.

California illustrates the disclosure problem. Reporting by CalMatters found that companies frequently do not disclose detailed facility-level water use, even when projects are proposed in water-stressed areas. Without consistent reporting, residents may hear that a project will use "less water than agriculture" or "about as much as a golf course," but have little basis for evaluating the claim.

The strongest argument against current data center development is not that society should stop using cloud computing or AI. It's that the industry is moving faster than public infrastructure and local planning systems can adapt.

Electricity is the central concern. Data centers can arrive as enormous new loads in regions that have experienced little demand growth for decades. Utilities must add generation, transmission, substations and distribution equipment. Those investments take years, while technology companies may want power within months.

Opponents worry that households will subsidize those upgrades through higher rates. They also fear that utilities will extend the lives of coal and natural gas plants, build new fossil-fuel generation or delay climate targets because renewable resources and transmission cannot be added quickly enough.

Those concerns are not hypothetical, but neither is the impact on the rate settled. A 2026 working paper from researchers at the Electric Power Research Institute and Watershed, examining U.S. electricity prices from 2015 through 2024, found that data center growth had modestly reduced average retail rates during that period, possibly because large, steady loads spread fixed grid costs across more electricity sales. The researchers warned, however, that future supply constraints could reverse that effect.

In other words, data centers may lower rates where grids have spare capacity and raise them where new infrastructure is expensive, or electricity supplies are tight.

Water produces a similar split. A facility using recycled water or closed-loop cooling in a water-rich region may present a manageable burden. The same facility using potable water during summer peaks in an arid region could threaten a public system's reserve capacity.

Noise is more local and less abstract. Cooling fans, chillers, transformers, and backup generators can produce a persistent mechanical hum. A highway quiets at night. A data center may not. Residents living near existing facilities have complained about low-frequency noise that is difficult to block and hard for conventional zoning codes to address.

Land use also matters. Data centers can occupy hundreds of acres, often near transmission corridors and population centers. They compete with housing, farms, warehouses, parks and other industrial development. Because they may employ relatively few people after construction, critics question whether they are the best use of scarce industrial land.

Then there are tax incentives. States and localities have offered data center developers exemptions from sales, property or equipment taxes. Supporters say those incentives attract investment that would otherwise go elsewhere. Critics say states are bidding against one another to subsidize some of the richest companies in the world.

Hochul said she would pursue legislation repealing New York's sales-tax exemptions for the largest data centers. That may prove as consequential as the temporary pause itself.

The argument in favor begins with a blunt fact: the modern economy already depends on data centers.

Hospitals, banks, manufacturers, government agencies, universities, retailers and small businesses use cloud infrastructure. Streaming video, online commerce, cybersecurity, communications, scientific computing and AI all require physical servers somewhere.

AI adds a new layer of demand. Training large models requires enormous clusters of specialized processors. Serving those models, known as inference, requires computing power every time a user enters a prompt, generates an image, writes code, translates a document or assigns a task to an AI agent.

Agentic systems can make the demand even larger because a single user request may trigger dozens of model calls, searches, database queries, tool operations, and verification steps.

Supporters therefore frame data centers as infrastructure rather than mere real estate. Countries that cannot provide electricity and computing capacity may become dependent on foreign AI systems, lose technology investment and weaken their position in strategically important industries.

The facilities also bring genuine economic benefits. Construction can employ large numbers of electricians, engineers, equipment operators, pipefitters and other skilled workers. Campuses can generate substantial property-tax revenue, even if their permanent workforces are smaller than those of traditional factories. They also purchase services from utilities, contractors, equipment suppliers, security companies and local businesses.

New demand can accelerate investment in generation and transmission that the grid needs anyway. Major technology companies have become some of the world's largest corporate purchasers of renewable energy. They are also investing in geothermal energy, batteries, advanced nuclear reactors and the restart or continued operation of existing nuclear plants.

Data centers can also operate more flexibly than many industrial facilities. Some computing workloads can be shifted to another time or location when the grid is under stress. Research suggests that even moderate workload flexibility could reduce transmission congestion and renewable-energy curtailment, although implementing such systems at commercial scale requires coordination among operators, utilities and customers.

The pro-development argument is not that data centers have no cost. It's that refusing to build them does not eliminate demand. It sends the investment, tax revenue, and computing capacity elsewhere.

Trump's "liquid gold" description captures that view, even if it overlooks some of the local trade-offs.

New York is the first state to impose a statewide moratorium, but it's not acting in isolation.

Lawmakers in 15 states have introduced legislation that would halt or restrict data center development, according to the National Conference of State Legislatures. Meanwhile, cities, counties and townships have adopted local moratoriums to rewrite zoning rules governing noise, setbacks, water use, emergency generators and electrical infrastructure.

Heatmap News found that at least 20 proposed projects, representing more than $41.7 billion in planned investment, were canceled following local opposition during the first quarter of 2026, a record for a single quarter.

The opposition does not follow a neat partisan pattern. Rural conservatives object to land seizures, transmission corridors, subsidies and threats to property values. Environmental groups focus on water, emissions and fossil-fuel generation. Labor unions support construction jobs but may oppose blanket moratoriums. Consumer advocates worry about electricity rates. Local officials want tax revenue, but they also want leverage.

That mixture may make the backlash more durable than a conventional left-versus-right policy fight.

New York's approach attempts to occupy the middle. The moratorium is temporary. It does not declare data centers inherently undesirable. It gives regulators one year to develop statewide environmental standards, cost-allocation rules and a community-benefits framework intended to help localities negotiate compensation and infrastructure commitments.

That could mean requiring developers to pay directly for grid upgrades, disclose peak water use, use recycled water where possible, meet noise standards, fund community projects, provide emergency power plans, or demonstrate that projects will not undermine climate targets.

Those requirements would make projects more expensive. Supporters would call that accountability. Developers might call it a reason to build somewhere else.

Both could be right.

The data center backlash is often portrayed as a collision between technological progress and public fear. That framing is too simple.

Many opponents use cloud services and AI. They are not rejecting computation. They are asking why their town should provide the land, water, electricity, tax breaks and political consent required to make somebody else's digital economy possible.

The industry has a strong answer. AI, cloud computing and digital services have become essential infrastructure. Building too slowly could constrain innovation, weaken economic competitiveness and leave power-hungry systems dependent on aging grids that need investment anyway.

But communities have an equally strong response. Essential infrastructure still requires rules. Airports face noise limits. Factories face emissions permits. Utilities face rate regulation. Stadiums negotiate public financing and community benefits. Golf courses operate under water restrictions. Data centers are not less physical simply because their product is digital.

Hochul's moratorium may eventually be remembered as a temporary political gesture. It may also mark the moment when the AI infrastructure boom lost its assumption of automatic consent.

The next phase will not be decided solely by who can build the largest model, buy the most GPUs or raise the most capital. It will also be decided at utility commissions, zoning boards, town halls and kitchen tables.

AI needs data centers.

The question is whether the companies building them can convince communities that they need AI's infrastructure, too.