AI Data Center Boom: Public Opposition, Costs & Financial Risks
The United States currently operates approximately 65 gigawatts (GW) of data centers, with an additional 413 GW in the planning stages. A gigawatt can power roughly 800,000 American homes, meaning the planned expansion is equivalent to connecting 330 million new homes in a country with 135 million households. Each gigawatt requires a power station, and the costs associated with this massive infrastructure buildout are substantial.
The artificial intelligence (AI) industry, which drives much of this demand, faces public skepticism. Despite previous warnings from AI developers about job displacement or even existential threats, they now seek to build these power-intensive facilities in communities, often requesting tax exemptions. A recent poll indicates that three out of four Americans oppose data centers in their communities, with public sentiment being more negative than for nuclear power stations or coal plants. This creates a dilemma for politicians who have historically competed to attract these projects for tax revenue and construction jobs, while voters express strong opposition.
Public Concerns: Water, Noise, and Electricity
Three primary concerns repeatedly arise regarding data center construction:
Water Usage
The perception that data centers drain vast water resources is largely unfounded. A mid-sized data center uses about as much water as two golf courses, and agricultural irrigation across the country consumes 74 times more water than data centers.
Noise Pollution
Noise is a significant issue. Data centers emit a low-frequency rumble that travels long distances and penetrates walls, often described by neighbors as a physical sensation rather than just a sound. Many viral videos of "screaming" data centers, however, are actually crypto mines, which consist of thousands of fan-cooled computers, each as loud as a vacuum cleaner.
Research by Ben Jordan has explored infrasound, frequencies too low for humans to hear, emitted by data centers and crypto mines. A pilot study found that exposure to infrasound led to increased discomfort and dizziness in participants. While AI data centers are generally quieter than crypto mines, a notable exception in Memphis involves a facility where the power supply is the loudest component.
Electricity Consumption
Electricity is the most substantial concern. Data centers consume enormous amounts of power, and their impact on utility bills is complex. The Electric Power Research Institute (EPRI), funded by the power industry, found that data centers built between 2019 and 2024 in typical states made power about 6% cheaper by helping cover fixed grid costs in areas with spare capacity. However, much of this spare capacity has been utilized. In the PJM grid (Illinois to New Jersey), the Independent Market Monitor estimates that planned data centers will contribute $6.3 billion to future capacity costs, arguing that data centers are not fully covering these expenses. Both claims can be simultaneously true, leading to public confusion and dissatisfaction.
Industry Attempts to Win Over Communities
The tech industry's primary strategy to gain community acceptance has been financial incentives. A University of Virginia poll found that residents would, on average, support a local data center for a payment of $10,000. The Economist calculated that paying every resident in a median-sized American county would cost around $240 million, which is considered "pocket change" for the tech industry.
Examples of these incentives include: * Richland Parish, Louisiana: A Meta data center led to $51,000 bonuses for many local teachers, exceeding their annual salaries.
Despite these payments, public opposition persists, partly due to how these deals are structured. Developers often operate through front companies with code names (e.g., "Project Canoli"), and local officials are frequently asked to sign Non-Disclosure Agreements (NDAs), preventing residents from knowing about projects until they appear on council agendas.
Controversial Deal-Making
- Perry Village, Ohio: The mayor and council members of this 1,600-person village signed NDAs with a California developer before negotiations. The planned data center would cover a sixth of the village and border a school baseball field. When residents forced a referendum, the council rescinded the zoning, then reinstated it under emergency provisions to prevent another vote, citing financial ruin for the village if the data center wasn't built.
- Saline Township, Michigan: After the board voted 4-1 against rezoning land for a data center, the developer sued the township for exclusionary zoning. The town settled, receiving several million dollars for the fire department and schools. This effectively means overruling a town's "no" vote costs millions, which may be a necessary investment given the fire department's new responsibility for a large building full of computers and backup batteries.
These tactics, involving secret documents, contracts that preclude referendums, and lawsuits, make the financial incentives appear less like community benefits and more like the cost of overriding local democracy.
The "Washington Style Problem"
Ezra Klein observed that Silicon Valley's worldview is shaped by solving "impossible" problems in code, while Washington's is defined by "possible" problems proving impossible to solve. The AI industry, having achieved seemingly impossible feats in software, now struggles with the "Washington style problem" of building physical infrastructure, which requires persuasion and navigating complex zoning boards rather than just functional code.
Elon Musk's Approach in Memphis
Elon Musk's AI company, XAI, faced similar challenges in building its supercomputer, Colossus 2, in Memphis, Tennessee. To circumvent the years-long process of building new power stations and upgrading the grid, XAI brought its own power plant across the state line to South Haven, Mississippi.
- Portable Power Plant: XAI deployed a fleet of gas turbines, initially 27, later growing to 69, to generate electricity. Mississippi regulations allow portable equipment to operate for up to a year without an air permit. Since the turbines arrived on trucks, they were legally considered "traffic." The newest units are permitted until July 2027.
- Community Impact: South Haven residents are less impressed. An acoustics firm found low-frequency noise near the plant to be 4 to 16 times louder than background levels in a quiet neighborhood, described as a constant "plane taking off but never leaving."
- Environmental Concerns: Portable turbines are less efficient and cleaner than standard gas plants. The NAACP's lawsuit alleges XAI's fleet could produce 26 times the smog-forming nitrogen oxides of a nearby TVA gas plant.
- Legal Battles: Residents filed a class-action lawsuit over lost sleep and falling property values, and the NAACP sued, arguing it's an unpermitted power plant. An alderman noted, "Are they unpermitted? Yeah, but they're not illegal."
- National Security Intervention: The Justice Department and Pentagon intervened on XAI's behalf, with the Justice Department arguing that keeping Grok (XAI's AI) running is a matter of "paramount national security."
- Memphis Benefits: Memphis secured an ordinance directing 25% of XAI's property taxes, up to $100 million, to surrounding neighborhoods. XAI also promised a water recycling plant, but construction stalled, leaving it half-built with an estimated $100 million needed to finish.
Interestingly, the supercomputer built to fact-check tweets is now being rented out to competitors, generating revenue. SpaceX, Musk's other company, spent $19.6 billion last quarter on chips, data centers, and related expenses, with spending expected to double this quarter. SpaceX now has more computing capacity than Coreweave, a major AI chip rental company.
The Financial Risks of the AI Buildout
The AI buildout is incredibly expensive, with tech companies borrowing "a cajillion dollars" (Bloomberg's Matt Levine's estimate) to finance it. A significant concern is the collateral for these loans: the computer chips themselves. Nvidia constantly releases newer, faster chips, meaning the collateral quickly becomes obsolete, while the buildings are designed to last decades. Lenders are being asked to provide long-term money against rapidly depreciating assets.
- Bond Market Concerns: In April, data center developer QTS issued $4.6 billion in bonds, initially bought eagerly 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 a new tenant for a massive, empty building. By August, QTS had to offer a 7.2% interest rate for new bonds, comparable to junk-rated companies.
- Coreweave's High-Yield Debt: Coreweave, a specialist cloud company, heavily relies on high-yield debt, much of it backed by its chips.
- Credit Default Swaps (CDS): The cost of insuring Coreweave's debt through credit default swaps is exceptionally high. Insuring $10 million of Coreweave debt for five years costs $1.95 million upfront (almost 20%) plus $125,000 quarterly, totaling about 11% of the insured amount annually. This high upfront cost reflects a market expectation of a greater than 60% chance of Coreweave defaulting within five years.
- Systematic Risk: Matt Levine highlights a structural problem: the typical sellers of AI credit risk insurance (banks, hedge funds, insurance companies) already hold significant AI data center risk and are unwilling to take on more without substantial compensation. This risk is largely systematic; if demand for AI computing falls short of projections, all data centers face the same problem simultaneously, a risk that cannot be diversified away. Five years ago, a five-year CDS on Oracle cost 0.33% annually; today, with Oracle heavily invested in AI, it costs 3.3%.
Who Bears the Risk?
The "smart money" is charging exorbitant rates to take on AI debt risk, raising the question of who is holding these bonds. Robert Armstrong of the Financial Times notes that an infrastructure buildout of this scale requires the investment-grade bond market. Achieving an investment-grade rating, which Coreweave is pursuing, allows pension funds, annuities, and life insurance companies to purchase the debt. This means that individuals with pensions, annuities, or life insurance policies are likely financing the very data centers they might be protesting.
Furthermore, the six hyperscalers and major chipmakers now account for about 20% of the global stock market. Therefore, anyone owning an index fund also holds a significant stake in these companies. While protesters might be part-owners through their pensions, those living next to data centers bear most of the noise and receive little of the upside, potentially sharing in the losses if the boom goes bust.
Lessons from Past Booms
The Midwest has experience with failed infrastructure promises. * Foxconn in Mount Pleasant, Wisconsin (2017): Foxconn promised 13,000 high-paying jobs, leading the village to invest hundreds of millions in infrastructure and tax incentives. Foxconn largely reneged, leaving the village with about 1,000 jobs and the infrastructure bill. The cleared land now hosts a Microsoft data center, where neighbors are suing over noise. * SolarCity in Buffalo, New York: New York spent nearly a billion dollars building a factory for SolarCity (later acquired by Tesla), expecting thousands of jobs. The state controller found only 54 cents of economic benefit for every dollar spent.
Today, utilities are upgrading grids for data centers, often requiring long contracts or exit fees to cover costs. However, if expected demand doesn't materialize, utilities will recover fixed costs from fewer customers, meaning remaining residents will pay for the transmission lines.
OpenAI once suggested the U.S. government should be the "insurer of last resort" for AI credit risk, a suggestion deemed "annoying" but understandable given the market-wide nature of the risk. The Justice Department's intervention on XAI's behalf further illustrates government involvement in protecting these projects.
Arguments for the AI Buildout
Despite the challenges, there are strong arguments for the AI buildout: * Job Creation: Data centers provide well-paid work for construction and electrical union members. A drainage worker in Ohio, for example, reported earning $24,000 a year at age 23, comparable to a junior investment banker. * Geopolitical Advantage: Making it harder to build AI infrastructure in the U.S. could cede advantage to countries like China, which can more easily direct such projects. * Demand Exceeds Supply: Unlike classic speculative bubbles, current demand for AI computing power exceeds supply. AI labs are rationing what they sell because they cannot acquire enough computing power. However, this demand is largely for free versions of chatbots, meaning companies are spending far more than they earn. The true demand at full cost is unknown.
Local moratoriums are unlikely to significantly slow down the buildout. Princeton professor Arvin Narayan calculated that an average one-year state-level moratorium would only delay AI capability growth by a matter of hours, as projects simply relocate to states without bans. Ireland, which saw data centers consume over a fifth of its electricity before the AI boom, effectively stopped connecting new ones in Dublin. This led to new rules requiring developers to bring their own power and buy renewables, demonstrating the industry's adaptability.
The AI buildout is projected to cost about 3.6% of America's GDP annually until 2032, a scale comparable to the railroad boom, the interstate highway system, and most of America's electrification combined.
Historical Parallels: The Railroad Bubble of 1873
Microsoft CEO Satya Nadella recommended "The Railroad Bubble of 1873" by Lea Quad Ahmed, a book that explores a historical period of massive infrastructure investment. Ahmed believes Nadella may have drawn the wrong lesson, noting that individual companies can act sensibly, yet the combined effect can still be wild overbuilding. Companies are not fully accounting for what all other tech companies are doing simultaneously.
During the 1873 railroad boom, American towns and counties borrowed heavily to buy railroad shares. When the panic hit, about a billion dollars of railroad bonds defaulted, leading to local government defaults and states rewriting constitutions to limit borrowing. Midwestern farmers, angered by railroad rates, formed the Granger Movement, which eventually led to the Interstate Commerce Commission, America's first federal regulator. The railroads were built, but many who paid for the buildout never recouped their investments.
Today, towns face a similar situation, but the exposure is more hidden, primarily through pension funds (owning shares and debt) and power bills (covering new grid costs if data centers leave). While the median price for a resident to accept a data center is $10,000, no one has asked what they would charge to take on the risk of the entire enterprise failing, a risk many already bear through their pensions.
Takeaways
- The United States plans to add 413 GW of data‑center capacity, enough to power roughly 330 million homes, requiring massive new power stations and huge infrastructure costs.
- Three‑quarters of Americans oppose data centers in their neighborhoods, citing concerns about water use, low‑frequency noise and the massive electricity demand they create.
- Tech firms offer cash incentives, but they often hide projects behind NDAs, secret code names and lawsuits that override local votes, making payments seem like a price for democracy.
- Bond issuances and high‑yield debt backed by rapidly depreciating AI chips carry soaring CDS spreads, exposing pension funds and other investors to systemic risk from a potential AI‑data‑center bust.
- Historical parallels to the 1873 railroad boom suggest the AI data‑center buildout could become a hidden, over‑invested infrastructure bubble whose costs fall on taxpayers, utilities and nearby residents.
Frequently Asked Questions
Why do some studies say AI data centers make electricity cheaper while others claim they add billions to grid capacity costs?
Some analyses find AI data centers lower electricity rates because they consume idle grid capacity, spreading fixed transmission costs across more load, while other estimates calculate that when that spare capacity is fully utilized the centers add billions in future capacity expenses. The two findings reflect different assumptions about how much unused capacity remains and how costs are allocated.
How does Elon Musk’s XAI portable power plant in Mississippi avoid traditional air‑permit requirements?
Elon Musk’s XAI sidestepped conventional air‑permit rules by shipping dozens of gas‑turbine generators on trucks and classifying them as portable traffic, which Mississippi law allows to operate for up to a year without a permit. This legal loophole lets the fleet run while the company argues the temporary setup is essential for national‑security AI workloads.
Who is Patrick Boyle on YouTube?
Patrick Boyle is a YouTube channel that publishes videos on a range of topics. Browse more summaries from this channel below.
Does this page include the full transcript of the video?
Yes, the full transcript for this video is available on this page. Click 'Show transcript' in the sidebar to read it.
of who is holding these bonds. Robert Armstrong of the Financial Times notes that an infrastructure buildout of this scale requires the investment-grade bond market. Achieving an investment-grade rating, which Coreweave is pursuing, allows pension funds, annuities, and life insurance companies to purchase the debt. This means that individuals with pensions, annuities, or life insurance policies are likely financing the very dat
centers they might be protesting.
Helpful resources related to this video
If you want to practice or explore the concepts discussed in the video, these commonly used tools may help.
Links may be affiliate links. We only include resources that are genuinely relevant to the topic.