Who Really Pays for AI Data Centers?

xAI

Across the shows: 2 speakers in other episodes, 1 bullish and 0 bearish calls, past 30 days

• Patrick Boyle · Patrick Boyle On FinanceBearish

  xAI's portable turbine fleet exploits a permit loophole, imposing noise and smog on South Haven residents.

  Boyle describes xAI bypassing normal power plant permitting by running gas turbines across the state line in Mississippi to supply its Memphis data center. Mississippi lets portable equipment run for up to a year without an air permit, so the turbines count as traffic. The company said it ran 27 turbines in January, and regulators counted 69 by July. An acoustics firm found low-frequency noise could be four to 16 times louder than background. The NAACP lawsuit claims the fleet could emit about 26 times the nitrogen oxides of the nearby TVA gas plant. Boyle concludes that the operation is unpermitted, even if not illegal, and that the community bears the costs.

  “Now, normally building a power plant that rivals the largest in the state involves things like air permits and public hearings, but Mississippi allows portable equipment to run for up to a year without an air permit.”at 12:42

QTS Realty Trust

• Patrick Boyle · Patrick Boyle On FinanceBearish

  QTS's data center bonds repriced from 5.7% to 7.2% as lease-expiry refinancing risk surfaced.

  QTS issued $4.6 billion of bonds in April at a 5.7% yield, mainly because Microsoft was the expected tenant. Investors then noticed the bonds would not be fully repaid before Microsoft's initial lease ends. That leaves QTS needing a new tenant for a very large building if Microsoft leaves. When QTS returned to the market in August, investors demanded 7.2%, about what junk-free top-rated companies pay. Boyle presents the rate increase as the market pricing in that refinancing and tenant risk.

  “mostly because Microsoft was going to be the tenant, but then investors read the terms and conditions more carefully and realized that the bonds wouldn't be fully paid off before Microsoft's initial lease runs out”at 17:11

CoreWeave

Across the shows: 44 speakers in other episodes, 11 bullish and 40 bearish calls, past 30 days

• Patrick Boyle · Patrick Boyle On FinanceBearish

  CoreWeave's chip-backed high-yield debt looks dangerously risky, with insurance prices signalling default fears.

  Boyle argues CoreWeave's data centers rely heavily on high-yield debt, much of it backed by its own chips, which become obsolete quickly. He cites the credit default swap market: insuring $10 million of its debt for five years costs about $1.95 million upfront plus $125,000 a quarter, roughly 11% a year. Bloomberg reports the swaps market prices a more than 60% chance of default within five years. He dismisses the co-founder's claim that insurance capital is unlimited, noting that insurers are reluctant to take on AI data center risk unless paid a large sum. Boyle regards the company's plan to win investment-grade ratings as an attempt to attract a different pool of buyers.

  “Things are even worse at CoreWeave, the specialist cloud company I mentioned earlier, which relies heavily on high yield debt to pay for its data centers, a lot of it backed by its chips.”at 17:45

Oracle

Across the shows: 16 speakers in other episodes, 0 bullish and 17 bearish calls, past 30 days

• Patrick Boyle · Patrick Boyle On FinanceBearish

  Oracle's default insurance cost rose roughly tenfold, signalling rising credit risk from its AI buildout.

  Boyle cites Matt Levine's observation that five-year credit default swaps on Oracle cost about 0.33% a year five years ago and about 3.3% today. He attributes the rise to Oracle being strapped into the AI buildout. Boyle presents this as evidence that the market now sees Oracle's credit risk as substantially higher because of its heavy AI spending. The argument rests on the change in CDS pricing rather than on Oracle's own financials.

  “Five years ago, Levine points out you could buy five year CDS on Oracle for about 0.33% a year. Today, with Oracle strapped into the AI build out, the same insurance costs about 3.3%.”at 20:45

AI infrastructure & data centers

• Patrick Boyle · Patrick Boyle On Finance

  Data center water concerns are overblown; farm irrigation uses far more water than data centers.

  Boyle says the water argument against data centers mostly falls apart when examined. Citing The Economist, he says a mid-sized data center uses about as much water as two golf courses. Crop irrigation across the country uses 74 times as much water as data centers. He suggests that if water were the real concern, protesters would also target country clubs. The point is that data centers are a small share of water demand.

  “a mid sized data center uses about as much water as two golf courses do and across the country crop irrigation uses 74 times as much water as data centers do”at 3:15

• Patrick Boyle · Patrick Boyle On Finance

  Data center costs are being pushed onto neighbours, ratepayers and pension savers, not the builders.

  Boyle argues the public, not the tech companies, carries the cost of AI data centers. Residents face noise and lost property values, while a few towns and counties receive payments or settlements that are small against the scale of the projects. Pension funds, which he says are often his own, hold the shares and debt of the builders, so savers finance the buildout while also paying through power bills. If the boom fails, grid costs and losses fall on the same households. He concludes the neighbours own the least of the project but absorb most of the noise and much of the downside.

  “The bad news is that the people who live next door own the least of it, they get most of the noise and very little of the upside, and if it all goes wrong, they will probably get a share of the losses too.”at 22:42

• Patrick Boyle · Patrick Boyle On Finance

  Local data center moratoriums won't slow AI much; projects just relocate to permissive states.

  Boyle cites Princeton professor Arvind Narayan, who calculated that an average one-year state-level moratorium would slow AI capability growth by only hours. He concludes the projects do not stop; they move to a state that has not banned them. Ireland shows the same pattern: the grid operator stopped connecting new Dublin data centers, but developers adapted by bringing their own power and buying renewables. The build-out continued. Boyle therefore expects local moratoriums to have little effect on the overall pace of AI infrastructure.

  “This is also why the local moratoriums probably won't slow things down much.”at 27:29

Power demand & the grid

• Patrick Boyle · Patrick Boyle On Finance

  Data centers aren't paying their own grid costs; ratepayers absorb future PJM capacity charges.

  Boyle reports that EPRI found data centers built from 2019 to 2024 lowered power costs by about 6% in the typical state, but only because they used spare grid capacity that covered fixed costs. That spare capacity is now largely used up. The PJM Independent Market Monitor estimates about $6.3 billion of future capacity costs come from planned data centers, and he says the data centers are not paying those costs themselves. Boyle adds that if expected demand never arrives, utilities recover fixed costs from fewer customers, so the remaining customers pay for transmission lines. He treats both the earlier benefit and the later burden as true.

  “...the Independent Market Monitor estimates that about $6,300,000,000 of future capacity costs come from planned data center costs”at 5:56

AI chips & semiconductors

• Patrick Boyle · Patrick Boyle On Finance

  AI chips will become obsolete far faster than the buildings housing them, mismatching long-term loan collateral.

  Boyle argues that much of the debt financing AI data centers is secured against the computer chips. Nvidia keeps releasing faster chips, so the chips inside will be obsolete long before the buildings, which last decades. Lenders are therefore lending long-term money against an asset that depreciates quickly. Boyle cites Carlos Mendez, who compared it to financing a fax machine just before email was invented. He presents this as a structural weakness in the financing.

  “A lot of this debt is secured against the computer chips themselves, and Nvidia keeps releasing better and faster ones. The buildings will last for decades, but it is pretty much guaranteed that the chips inside them will be obsolete long before that.”at 16:24

US state & local policy

• Patrick Boyle · Patrick Boyle On Finance

  Big local incentive deals for factories and data centers often deliver far fewer jobs than promised.

  Boyle gives two examples of large local incentive deals that fell short. In Mount Pleasant, Wisconsin, a village of 28,000 spent hundreds of millions on roads, water lines and tax incentives for Foxconn, which promised 13,000 jobs. The result was about a thousand jobs and the infrastructure bill. New York State spent almost a billion dollars on a Buffalo factory for SolarCity, and its controller later found about 54 cents of economic benefit per dollar spent. Boyle uses these cases to suggest that the data center subsidies now being offered carry similar risks.

  “Foxconn then more or less changed its mind, and the village ended up with about a thousand jobs and the bill for the infrastructure.”at 23:26

AI bubble & tech valuations

• Patrick Boyle · Patrick Boyle On Finance

  AI buildout risks railroad-style overbuilding; current demand exists only because AI is nearly free.

  Boyle weighs the case that AI is not a bubble. Labs are rationing compute, which suggests demand exceeds supply. He argues that this demand is at today's prices, and for most users the price is zero, while the companies spend far more than they earn. Nobody knows how much demand would remain if everyone paid full cost. He draws on a railroad analogy: each company can be sensible alone while the combined result is wild overbuilding, and six hyperscalers plus the big chipmakers make up about 20% of world stock market value. He cites an estimate that the buildout will cost 3.6% of US GDP a year until 2032. Boyle's conclusion is that the risk of overbuilding is real.

  “The catch is that this is demand at today's prices, and for most people today's price is zero.”at 27:00