Who Really Pays for AI Data Centers?

Who Ultimately Pays for the AI Data-Center Boom?

The video examines who bears the costs and risks of the rapid expansion of AI data centers. The United States has about 65 gigawatts of data-center capacity, with another 413 gigawatts planned—an expansion that would require enormous amounts of electricity and new power infrastructure.

Local opposition is growing as residents worry about the effects on their communities. But the financial consequences may reach far beyond the places where data centers are built.

Local Impacts and Electricity Demand

Residents’ concerns vary in how well they are understood. Nationally, data centers use much less water than crop irrigation, though their effects can still matter locally. Noise is a more tangible concern: low-frequency hum can travel through walls, and crypto-mining sites—with large numbers of fan-cooled machines—are sometimes mistaken for AI data centers.

A small study by science YouTuber Ben Jordan found more discomfort and dizziness among people unknowingly exposed to recorded infrasound. Its effects remain poorly understood, however, and require further research.

Electricity is the most substantial concern. One industry-funded study found that data centers helped lower average electricity costs in some states by contributing to existing grid capacity. But that spare capacity is diminishing. In the PJM region, a market monitor estimated that billions of dollars in future capacity spending were associated with planned data centers.

Both claims can be true: data centers may have helped pay for grids when capacity was underused, while new demand now risks driving expensive upgrades and higher bills.

Community Benefits and Local Opposition

The industry has tried to win support through payments and other local benefits. A University of Virginia survey found that residents needed a median payment of $10,000 to support a local data center. Paying that amount to every resident of an average county would cost roughly $240 million. Meta’s Louisiana project, for example, funded teacher bonuses of up to $51,000.

Money does not always overcome distrust. Investigations cited in the video describe developers using shell companies, asking officials to sign nondisclosure agreements, and pursuing projects through zoning processes that residents say gave them little meaningful opportunity to object.

In Saline, Michigan, a developer sued after the council rejected a zoning change. The town ultimately settled in exchange for funding for a fire station and schools.

The xAI Project Near Memphis

Elon Musk’s xAI offers a direct example of the trade-offs. To power its Memphis-area supercomputer, the company installed a large fleet of gas turbines in nearby Mississippi, taking advantage of a rule that allows portable equipment to operate temporarily without a permit.

Residents have complained of persistent noise, and an acoustic study found low-frequency noise substantially above the background level. The NAACP also alleged that the turbines produce far more smog-forming nitrogen oxides than a nearby conventional gas plant.

Residents and the NAACP sued. The Justice Department asked a court to dismiss the case, citing national-security interests. Memphis negotiated a commitment to direct up to $100 million in property-tax revenue toward development in the surrounding area, while a promised water-recycling facility remained unfinished.

Financing Risks

The video also examines how the AI buildout is financed. Data-center loans are often secured by computer chips that may become obsolete much faster than the buildings that house them. Bond investors have also begun to question whether facilities will retain tenants over the long term. QTS bonds initially sold at a 5.7% yield, but later borrowing was priced at 7.2%—near levels associated with junk-rated companies.

CoreWeave is presented as a particularly risky borrower. Insurance against a possible default on $10 million of its debt was quoted at $1.95 million upfront plus $125,000 quarterly, or about 11% of the insured amount per year. The swaps market was said to imply a greater-than-60% chance of default over five years.

Yet the company’s co-founder said CoreWeave hoped to obtain investment-grade ratings for much of its debt, which could attract buyers such as insurers. The concern is that a broad decline in AI demand would affect many data-center companies at once, making the risk systemic rather than easy to diversify away.

That debt can end up in pensions, annuities, and insurance portfolios. Major technology companies and chipmakers also make up a large share of global stock-market value, exposing index-fund investors to the buildout. Residents near a data center may therefore own a small indirect stake in it through their savings while bearing a disproportionate share of noise, local disruption, and potentially higher utility costs.

Lessons from Past Infrastructure Booms

The video compares the buildout with earlier infrastructure projects. Foxconn’s Wisconsin project and a publicly subsidized factory in Buffalo illustrate how communities can be left with costly infrastructure and fewer jobs or benefits than promised. Utilities face a similar risk: if expected data-center demand never arrives, remaining customers may have to pay for grid upgrades.

The video also draws on the 1873 railroad bubble, when local governments borrowed heavily to attract railroads and many ultimately defaulted. The railways were built and transformed the country, but many of those who financed them did not recover their money.

Arguments for Building

The video also presents arguments in favor of data-center construction. Projects create well-paid construction and trade jobs, and restricting American infrastructure could benefit China. Current demand for computing reportedly exceeds supply, but much of that demand comes from services users access for free or at prices that do not cover providers’ costs. It is therefore uncertain how much demand would remain if users had to pay the full price.

Local moratoriums may slow projects only slightly, since companies can move them to other states. Ireland’s experience suggests that restrictions can instead push developers to provide their own energy or meet other requirements.

Conclusion

The AI infrastructure boom may deliver lasting economic and technological benefits, but its costs and risks are widely distributed and often difficult for communities to see. Residents may bear local impacts, utility customers may cover unused grid investments, and pension or insurance holders may absorb financial losses if demand falls short.

The railroad analogy is not that the infrastructure will necessarily be useless. It is that useful infrastructure can still be built through overexpansion that leaves many of its funders worse off.

Presenters and Contributors Referenced

Patrick Boyle (presenter); Ben Jordan, Jasmine Sun, Ezra Klein, Matt Levine, Carlos Mendez, Toby Nangle, Brennan McBee, Robert Armstrong, David Wallace-Wells, Arvind Narayanan, Sven Neureiter, Satya Nadella, Liaquat Ahamed, and Derek Thompson.