Who Really Pays for AI Data Centers? Unpacking the Costs and Controversies

The United States currently has 65 gigawatts of data center capacity, according to energy data firm Clean View. However, plans are in motion to add an astonishing 413 gigawatts, with completion dates for the largest projects often listed as 'To Be Determined'. This planned expansion highlights a significant increase in demand for computational power, primarily driven by the growth of artificial intelligence.

To put this into perspective, one gigawatt of electricity is sufficient to power approximately 800,000 American homes. Therefore, the projected pipeline of data center expansion is equivalent to plugging in about 330 million new homes. This figure dramatically surpasses the current number of households in the U.S., which stands at around 135 million, indicating a substantial increase in electricity demand that requires new power stations and financial backing.

The rapid buildout of AI data centers faces significant public opposition, exacerbated by fears surrounding AI's impact on employment and humanity. Developers' proposals to build these facilities in residential areas, often seeking tax exemptions on components, have not been well-received. A recent poll indicates that three out of four Americans oppose data centers in their communities, positioning them as less favorable than even nuclear or coal power plants.

This strong public sentiment creates a dilemma for politicians. Governors from both major parties have actively competed to attract these projects for the tax revenue and construction jobs they promise. However, they are increasingly pressured by voters who are not keen on having data centers built near their homes. This tension has led to municipal zoning disputes, with citizens, often indirectly through their pension funds, unknowingly contributing to the very projects they protest.

Community opposition to data centers primarily revolves around three key concerns: water usage, noise pollution, and electricity consumption. While water consumption is frequently highlighted, its actual impact is often less significant than commonly perceived. For instance, a midsize data center uses roughly as much water as two golf courses, and nationally, crop irrigation consumes 74 times more water than data centers. This suggests that if water were the primary concern, protests would also target other high-water-use industries.

Noise, particularly low-frequency rumble, presents a more genuine problem, as it travels far and penetrates walls, causing residents to report feeling it as much as hearing it. Many viral videos depicting loud data centers are, however, often crypto mines, which use thousands of fan-cooled computers, each as loud as a vacuum cleaner. Research by Ben Jordan on infrasound, sound too low for humans to hear, indicates it can cause discomfort and dizziness. Electricity usage is the third main concern, with complex effects on local utility bills, as discussed in further detail.

In Memphis, a notable exception for noisy AI data centers, the loudest component of the operation is its power supply, rather than the data center itself.

Electricity is where complaints about data centers gain the most substance, as these facilities consume enormous amounts of power. The Electric Power Research Institute, primarily funded by the power industry, found that between 2019 and 2024, power in a typical state would have been approximately 6% more expensive without data centers. This benefit largely stemmed from areas with spare grid capacity, where data centers helped distribute fixed costs.

However, this dynamic is changing as much of that spare capacity, particularly in the PJM grid extending from Illinois to New Jersey, has been depleted. The Independent Market Monitor estimates that future capacity costs, specifically those attributed to planned data centers, will amount to about $6.3 billion. This monitor argues that data centers are not adequately covering their share of these capacity costs, a claim that can coexist with the earlier findings of price benefits.

The simultaneous truth of both claims—that data centers can reduce costs in some scenarios while adding substantial costs in others—demonstrates the complex and often contradictory economic impact of these facilities on utility bills and grid infrastructure, typically leaving local communities with unresolved concerns.

Tech companies have resorted to financial incentives to garner local support for data center construction. A University of Virginia poll revealed that the median payment residents would accept to support a local data center is $10,000. The Economist calculated that paying every resident in a median-sized American county would cost around $240 million, which it described as 'pocket change' for the tech industry, noting that voters appear quite receptive to such 'bribes'.

An example of this approach is seen in Richland Parish, Louisiana, where a Meta data center project provided substantial bonuses to local teachers, with many receiving $51,000—more than their annual salaries. A first-grade teacher described this influx of funds as feeling like 'something that happens to people in movies.' Despite these significant payouts, the effectiveness of these financial incentives in securing long-term community acceptance is questionable.

Data center developers frequently employ opaque tactics to secure local zoning approvals, often using front companies with innocuous code names like 'Project Canoli' or 'The Barn' to conceal their identities, as journalist Jasmine Sun reported in Wisconsin and Michigan. Local officials are commonly asked to sign non-disclosure agreements (NDAs), keeping residents uninformed about proposed projects until they appear on council agendas.

In Perry Village, Ohio, a community of 1,600, the mayor and key council members signed NDAs with a California developer even before negotiations commenced. The planned data center would have covered about a sixth of the village and bordered the school baseball field. When residents forced a referendum, the council rescinded the zoning, then reinstated it under emergency provisions at an evening meeting, preventing another vote. A county commissioner indicated that the choice was between building the data center or financially ruining the village.

Similarly, in Saline Township, Michigan, the board voted against rezoning land for a data center. The developer subsequently sued the township for exclusionary zoning. The town, with a few thousand residents, settled the lawsuit, agreeing to a few million dollars for its fire department and schools. These incidents highlight how developers leverage legal and financial pressure to overcome local opposition, often at the expense of community transparency and democratic processes.

Ezra Klein observed a fundamental difference in worldview between Silicon Valley and Washington D.C.: in Silicon Valley, impossible problems are often proven solvable, while in Washington, possible problems frequently prove impossible. This distinction highlights the tech industry's struggle to translate its code-based problem-solving approach to the physical realm of infrastructure and local zoning.

AI developers, adept at resolving complex coding challenges, find navigating local zoning boards much more arduous and time-consuming. The industry has largely sought to avoid the laborious process of persuading local zoning boards, indicating a disconnect between their fast-paced digital world and the slower, more deliberative processes of physical development.

Faced with lengthy delays in obtaining power for his xAI company's supercomputer, Colossus 2, in Memphis, Tennessee, Elon Musk implemented an innovative solution. xAI established a fleet of mobile gas turbines across the state line in South Haven, Mississippi. This setup allows the gas to burn in Mississippi, with the generated electricity being fed to the data center in Tennessee.

This approach exploits a regulatory loophole: Mississippi allows portable equipment to operate for up to a year without an air permit. The turbines, arriving on trucks, are legally considered 'traffic' because they have wheels, effectively bypassing the stringent permitting processes associated with permanent power stations. The number of these turbines rapidly increased from 27 in January to 69 by July, with permits for the newest units extending until July 2027, allowing a continuous temporary operation.

The residents of South Haven, Mississippi, are less enthusiastic about xAI's innovative power solution. An acoustics firm found that the low-frequency noise near the plant could be 4 to 16 times louder than the background levels in a quiet neighborhood, with residents describing it as sounding like a plane constantly taking off. These portable turbines are also less efficient and cleaner than standard gas plants, leading to significant environmental concerns.

The NAACP filed a lawsuit alleging that xAI's South Haven fleet could produce approximately 26 times the smog-forming nitrogen oxides of a nearby TVA gas plant. Residents have also initiated a class-action lawsuit, citing lost sleep and diminished property values. Although a South Haven alderman noted the turbines are 'unpermitted' but 'not illegal,' the Justice Department and the Pentagon intervened on xAI's behalf, arguing that keeping Grock running is a matter of paramount national security, asking the court to dismiss the suit.

Memphis has secured some benefits from the xAI deal, including an ordinance that redirects 25% of the property taxes from xAI's facilities, up to $100 million, to surrounding neighborhoods. The city mayor stated that xAI never requested a tax break. However, other promises have fallen short; xAI pledged a water recycling plant for cooling, but its construction stalled when it was half-built, requiring another $100 million to complete.

This $100 million completion cost is minor compared to the broader financial landscape of AI computing. Scientific American reported that this amount is less than three days of what Anthropic pays to rent computing power, which stands at $1.25 billion per month. This highlights the enormous financial scales involved in AI infrastructure. Furthermore, SpaceX, also owned by Elon Musk, is projected to roughly double its spending on chips and data centers this quarter, indicating massive ongoing investment in computing capacity.

Building supercomputers is immensely expensive, leading tech companies to borrow substantial capital, reportedly in the realm of 'a cajillion dollars,' to finance AI data centers. A significant concern for lenders is the collateral for these loans: computer chips. While the data center buildings themselves are expected to last for decades, the chips inside them are likely to become obsolete much sooner due to rapid technological advancements from companies like Nvidia.

Carlos Mendes of Crayill Capital likened current lending practices to 'financing a fax machine and then someone invented email,' emphasizing the rapid obsolescence risk. The bond market has begun to reflect this caution. In April, data center developer QTS issued $4.6 billion in bonds with a 5.7% yield, primarily due to Microsoft being a tenant. However, investors later realized the bonds wouldn't be fully paid off before Microsoft's initial lease expired, meaning QTS would need to find new tenants for large, empty buildings. By August, when QTS sought more funding, investors demanded a higher interest rate of 7.2%, typical for junk-rated companies.

Companies like CoreWeave, a specialized cloud company heavily reliant on high-yield debt backed by its chips, face even greater scrutiny. This increasing caution among lenders and investors highlights the significant financial vulnerabilities inherent in financing rapidly evolving AI infrastructure.

The financial market's concern over AI data center debt is evident in the pricing of credit default swaps (CDS). Insuring $10 million of CoreWeave's debt for five years demands an upfront payment of $1.95 million, nearly 20% of the insured amount, plus quarterly payments of $125,000. This totals an annual cost of about 11% of the insured amount, a significantly high premium.

This high upfront cost is a protective measure for CDS sellers, ensuring they are not left with substantial losses if the company defaults shortly after the agreement. Bloomberg reports that the swaps market is currently pricing in a greater than 60% probability that CoreWeave will default within the next five years, reflecting the market's perception of high default risk for companies in the AI data center sector.

Matt Lavine highlights a structural problem in the AI data center market: the risk is largely systematic and cannot be diversified away. Unlike idiosyncratic risks, such as a single company failing due to poor management, a downturn in overall AI computing demand would affect all data centers simultaneously. This means that if AI demand falls short of projections, all investments in this sector would face the exact same problem at the same time, making it impossible to mitigate risk through diversification.

Consequently, traditional insurers like banks, hedge funds, and insurance companies, which already hold significant AI data center risk, are hesitant to take on more without massive premiums. The increasing systematic risk is exemplified by Oracle, whose 5-year credit default swap cost rose from 0.33% to about 3.3% per year as its operations became more intertwined with the AI buildout.

Despite the high costs and risks, the sheer scale of AI infrastructure development means it relies heavily on the investment-grade bond market. Companies like CoreWeave are actively seeking investment-grade ratings to attract capital from large institutional investors such as pension funds, annuities, and life insurance companies. This implies that many individuals who possess pensions, annuities, or life insurance policies are inadvertently financing the very data centers they might be protesting.

Furthermore, with the six hyperscalers and major chip makers now accounting for about 20% of the global stock market, individuals holding index funds also own a significant stake in these companies. While protesters and indirect investors may benefit from potential AI boom gains, local residents living near data centers bear the brunt of noise, environmental impact, and other physical burdens, often receiving very little of the upside and potentially sharing in the losses if the boom goes bust.

The history of large-scale corporate infrastructure projects in the US offers cautionary tales about potential risks for local communities if the AI boom goes bust. In 2017, Foxconn promised 13,000 high-paying jobs in Mount Pleasant, Wisconsin, leading the village of 28,000 people to invest hundreds of millions in infrastructure and tax incentives. However, Foxconn largely changed its plans, leaving the village with only about a thousand jobs and the bill for the infrastructure. The cleared land is now home to a Microsoft data center, where neighbors are suing over noise.

Similarly, New York State spent nearly a billion dollars building a solar factory for SolarCity in Buffalo, expecting thousands of jobs. The state controller later determined that for every dollar spent, only about 54 cents of economic benefit was realized. Today, utilities are upgrading grids for data centers, often requiring long contracts or exit fees from these facilities. However, the Electric Power Research Institute notes that if expected demand doesn't materialize, utilities will recover fixed costs from fewer customers, meaning non-leaving residents will effectively pay for these upgrades.

Despite concerns, there are strong arguments supporting the AI infrastructure buildout. Data centers are generating numerous high-wage jobs for electrical and drainage workers. For example, a union official noted that members' hours have doubled in a decade, and a 23-year-old drainage worker in Ohio reported earning $24,000 a year. There is also a geopolitical argument: hindering domestic AI infrastructure development could grant a strategic advantage to countries like China, where leaders can site data centers without local opposition.

The most compelling argument against the current situation being a bubble is that demand for AI computing currently exceeds supply. As David Wallace Wells points out, this is not a classic speculative bubble where construction outpaces demand; AI labs are actually rationing their services due to insufficient computing power. However, this demand is largely at today's prices, which for most users, means zero cost. The actual demand if everyone had to pay the full cost remains unknown, leading to questions about the true sustainability of the current growth.

Matt Lavine notes that OpenAI once suggested the U.S. government should be the insurer of last resort for AI credit risk. He described that as an annoying suggestion at many levels, but added that he kind of gets it, since this is a market-wide risk and everyone is in it together. The government's posture is also visible in court: when xAI's turbines were challenged in court, the Justice Department sided with it.

Local moratoriums and regulations aimed at slowing data center growth are largely ineffective. Arvin Narayanan, a Princeton professor, calculated that a typical one-year state-level moratorium would only delay the growth of AI capabilities by a matter of hours. These projects do not stop; instead, they simply relocate to states or jurisdictions that have not yet implemented bans.

Ireland's experience provides a clear precedent. Before the AI boom, ordinary cloud data centers caused them to consume over a fifth of the country's electricity. A few years ago, Ireland's grid operator effectively halted new connections in Dublin. This did not stop the buildout but rather led to new rules requiring developers to provide their own power and purchase renewables. The industry simply adapted, illustrating the difficulty of curbing this rapid expansion through regional regulations.

The current AI infrastructure buildout is of a scale comparable to monumental historical projects. Economist Stijn Van Nieuwerburgh estimates that the AI expansion will cost approximately 3.6% of America's GDP annually until 2032, a figure equivalent to the combined costs of the railroad boom, the interstate highway system, and most of the country's electrification. This massive investment evokes parallels with past periods of rapid expansion and potential overbuilding.

Microsoft CEO Satya Nadella's interest in a history of the 1873 railroad bubble, a book described as 'the book to be read,' highlights concerns about potential overbuilding. Author Lea Quad Ahmed noted that Nadella might be drawing the wrong lesson, suggesting that while individual companies might act sensibly, their collective actions can lead to wild overbuilding. Ahmed told the Financial Times that companies are not adequately considering the cumulative effect of all tech companies pursuing the same expansion simultaneously.

The 1873 panic offers a historical parallel. During the railroad boom, American towns and counties borrowed heavily by selling bonds to buy shares in railroads they hoped would pass through their town. When the panic hit, around a billion dollars of railroad bonds defaulted, many local governments defaulted on their own debts, and states rewrote their constitutions to limit how much towns could borrow. Midwestern farmers angry about railroad rates organized into the Granger Movement, which eventually helped produce America's first federal regulator, the Interstate Commerce Commission. The railroads still got built, but the people who paid for the buildout never got their money back. Today the exposure is better hidden, flowing through pension funds that own the shares and debt, and through power bills that would cover the cost of new grid infrastructure if the data centers ever leave.