It Started: The AI Bank Run

Overview

The presenter argues that the rapid build-out of AI data centers is increasingly funded by private-credit lenders and complex borrowing structures. He warns that if investors seek withdrawals faster than funds can pay them, lending could freeze and financial stress could spread from data-center projects to technology companies, pension and insurance investments, and other parts of the financial system.

He frames this as a potential “AI bank run,” while emphasizing that the main risk is financing and liquidity—not necessarily the long-term usefulness of AI.

AI Infrastructure Spending and Financing

• Amazon, Microsoft, Meta, and Alphabet are said to plan roughly $725 billion in AI-related infrastructure spending this year, 77% above the prior year. The presenter says Wall Street expects spending to exceed $1 trillion next year.
• Morgan Stanley is cited as forecasting about $2.9 trillion in data-center spending by 2028. The presenter estimates that big tech’s own cash would cover roughly half, leaving about $1.5 trillion to borrow.
• Morgan Stanley is also cited as expecting private credit to provide approximately $800 billion of that financing gap.
• Private credit is described as non-bank funds lending directly to companies. The video says the sector expanded after post-2008 bank regulations constrained some bank lending, and argues that private-credit funds face lighter oversight than banks.

Data-Center Deals and Private-Credit Exposure

• Meta’s planned Hyperion data center in Louisiana is presented as an example. Meta reportedly owns 20% of a separate project company, while Blue Owl owns 80%. The project company borrowed about $27 billion through bonds due in 2049. The presenter says this debt is outside Meta’s own balance sheet.
• Blue Owl and similar funds are described as having attracted investors with returns of around 9% a year. The presenter distinguishes those past returns from withdrawal terms: investors could generally request withdrawals quarterly, but funds typically limited payouts to around 5% of fund assets.
• The video says Blue Owl permanently stopped withdrawals from one fund and sold $600 million of loans to raise cash. Its stock reportedly fell by as much as 9% in a day.
• Blackstone reportedly received record withdrawal requests from one fund. The firm and its employees contributed about $400 million to help meet them.
• Investors reportedly requested withdrawals equal to about 22% of a $36 billion Blue Owl fund, but received only 5%. The presenter says Ares, Apollo, and BlackRock also faced requests of roughly 9–12%. Across the funds discussed, about $15.6 billion was reportedly requested in one quarter.
• In a later example, Blackstone’s fund again received requests of about 10% but paid out 5%. The presenter describes this as the second consecutive quarter of restricted withdrawals.
• A UBS scenario is cited in which private-credit defaults could reach about 15%—roughly one in seven loans.

AI-Related Credit Concerns and Oracle

• The presenter says private-credit lenders previously favored software companies because of their recurring subscription revenue. He worries that AI tools could disrupt some software businesses’ ability to repay loans.
• Blue Owl reportedly sought to raise $4 billion for a data-center project whose planned tenant was CoreWeave, an AI-computing provider. The presenter says lenders judged the deal too risky, in contrast to earlier enthusiasm for financing AI data centers.
• Oracle is described as having made a large bet on renting AI computing capacity, including to OpenAI. Its long-term debt is said to have risen to about $125 billion. The presenter says S&P lowered Oracle’s credit rating to the lowest tier still considered investment grade, while the cost of insuring against default reached a record.
• Oracle’s Project Jupiter in New Mexico is described as a data center for OpenAI that would require power equivalent to about 1.8 million homes. Plans to use natural gas reportedly faced setbacks after state officials rejected a pipeline route.
• The presenter says Oracle sent a Blue Owl developer a force majeure letter that could allow it to delay rent for up to three years. Oracle reportedly said the project remained on schedule, and Blue Owl said nothing had changed.
• Some Oracle bonds are said to have traded at about 77 cents on the dollar, which the presenter interprets as a sign of heightened concern about repayment.

OpenAI’s Projected Spending

• Citing the Financial Times, the video says OpenAI expected about $840 billion in revenue and $856 billion in spending on computing power and infrastructure between 2026 and 2030—$16 billion more than projected revenue, before other expenses.
• The presenter says OpenAI could burn through another $278 billion in cash over that period.
• The financing chain is described as dependent on continual funding: OpenAI needs investors to pay Oracle; Oracle needs loans for infrastructure; and lenders need their investors not to withdraw.
• The presenter also notes that falling AI operating costs could benefit customers but pressure companies whose borrowing costs or long-term obligations remain fixed. Jeffreys is cited for AI running costs reaching their lowest level of the year in August; Siemens and PNC are mentioned as considering cheaper AI models.

Potential Spillovers and Historical Comparisons

• The video says buyers of bonds linked to Meta’s Hyperion project included PIMCO and BlackRock, which manage money for pension funds and insurers.
• The Bank for International Settlements is cited as saying AI-related lending rose from less than 1% to nearly 8% of private credit, and warning that a downturn could unwind faster than a conventional banking crisis.
• The presenter says the private-credit industry has lobbied to make such funds available in 401(k) plans.
• Historical comparisons include Britain’s 1840s railway boom and the early-2000s build-out of internet cables. The presenter’s point is that investors and lenders can lose heavily financing an infrastructure boom even if the infrastructure and technology later prove valuable.

Presenter’s Suggested Framework

The presenter says this is his opinion, not financial advice, and suggests:

• Check liquidity: Review pensions, ISAs, 401(k)s, and other holdings for terms such as “semi-liquid,” “interval fund,” or “private credit.” If money cannot be accessed quickly, he suggests keeping that exposure to a small share of the portfolio.
• Diversify: Avoid concentrating a portfolio solely in major technology companies or AI. Spread investments across industries and countries; he says this does not mean avoiding AI or big tech altogether.
• Monitor withdrawals: Check quarterly withdrawal requests at large private-credit funds. He treats shrinking queues as a sign of easing pressure and growing queues as a warning.
• Maintain cash capacity: He suggests having some cash available if market stress creates opportunities, noting that sell-offs can put otherwise attractive assets “on sale.”

Assets, Companies, Sectors, and Instruments Mentioned

• Companies and organizations: Amazon, Microsoft, Meta, Alphabet, Blue Owl, Blackstone, Ares, Apollo, BlackRock, Oracle, OpenAI, CoreWeave, Siemens, PNC, UBS, Morgan Stanley, S&P, PIMCO, the Bank for International Settlements, and Jeffreys.
• Sectors and themes: AI, data centers, cloud and AI computing, software, private credit, banking, pensions, and insurance.
• Instruments and investment products: Bonds, private-credit funds, stocks/shares, index ETFs, pensions, ISAs, 401(k) plans, and cash.
• Tickers: None are given in the subtitles.

Disclosures and Sponsorship

• The presenter explicitly says the discussion is not financial advice and that he is not a financial adviser.
• The presenter discloses that eToro sponsored a segment. He promotes an offer of a free share in one of nine companies after using his link and depositing £100 or the currency equivalent. The video also mentions eToro’s ready-made “smart portfolios.”

Presenter and Sources

• Presenter: Mark Tilbury.
• Sources cited in the video: Morgan Stanley, UBS, the Financial Times, the Bank for International Settlements, and S&P. Jeffreys is also cited regarding AI operating costs.