← All articles

Mikhail Polyakov

Banks are trying to resell the debt for AI data centres: what is reported and what a company that builds its work on cloud AI services should take from it

Level: IntermediateReading time: 5 min

Topics: Fintech, AI

A stack of plain banknotes and a cardboard model of a data-center building on a dark table beside balance scales with one pan tipped down

Short answer: according to a Startup Fortune report of 11 October 2026, banks that recently lent money for building data centres for AI are now looking for ways to shift part of that risk onto other investors. One example: a group of banks has been trying for more than six months to distribute $38 billion of debt on a project leased to Oracle. This is a third-hand report, and it mixes reported facts with the author's assessments. Below is a breakdown by part and what a company that buys cloud AI services should take from it.

What is reported

There is one source: a Startup Fortune article of 11 October 2026. It in turn refers to reports by the Financial Times and Bloomberg. I have not read those publications, so everything below is a retelling of a retelling, and the figures should be checked against the primary sources.

  • Which banks. According to Startup Fortune, citing the Financial Times, JPMorgan, Morgan Stanley, SMBC and MUFG are exploring significant risk transfers. These are deals in which a bank keeps a loan on its books and passes the riskiest part of it to outside investors.
  • Why. According to the text, the banks have come up against internal limits on how much AI-related exposure a single lender can hold.
  • An example with an amount. A group of banks including JPMorgan and MUFG, according to Bloomberg, has for more than six months been trying to distribute $38 billion of debt for building data centres in Texas and Wisconsin that are leased to Oracle. According to the text, some lenders are ready to sell pieces of the debt to non-bank buyers at a discount.
  • Market assessment. Matthew Moniot, co-head of credit risk sharing at Man Group, described the sizes of such deals as "out of scale to anything we've thought about, ever".

Which figures are given

  • Oracle. Five-year credit default swaps, that is, the price of insurance against the company failing to pay its debts, rose to around 151 basis points in early October. A basis point is one hundredth of a percent: 151 points means 1.51% a year of the insured amount. A level not seen since 2009, according to market data cited by Bloomberg. The rating agency S&P cut Oracle's rating in July to BBB-, one notch above junk. According to the text, the company is reportedly aiming to raise more than $40 billion of additional debt and equity this fiscal year.
  • Demand for bonds. An order book is a list of those who want to buy a bond issue, with the amounts. Such books for issues by Alphabet, Amazon, Meta, Microsoft and Oracle shrank from roughly five times the amount on offer in February to under two times by July (Bloomberg).
  • The price of money. The yield on 10-year US Treasury bonds reached 5.304% this month, the highest since 2002. The minutes of the Federal Reserve's September meeting, according to the text, named AI-related corporate borrowing as one contributing factor. Analysts at Capital Economics, as the article relays, think the main cause is inflation and government deficits, not tech companies' bond sales.
  • The difference in cost. Hyperscalers are the largest cloud companies, such as Microsoft and Google. According to CCIR data, paper linked to data centres leased by CoreWeave is priced at yields of 7.00% to 9.875%, and paper linked to data centres leased by hyperscalers at 5.70% to 7.875%. That is, the first has the higher yield.

Where the facts are and where the author's opinion is

The article contains many assessments that are not reports from banks or the media. For example, the author writes that banks "do not sell down debt they are confident in" and that the risk has ended up "at the banks that financed these structures". This is his interpretation. The source has no direct statements from banks about their reasons, apart from the reference to internal limits. The author also concludes that the assumption that credit remains available on reasonable terms is being tested in practice. This is his conclusion, not a measured indicator.

What a company should take from this

If your work depends on a cloud AI service, a single headline about a supplier's debts does not mean the service will stop. There is nothing of the kind in the source, nor any forecasts about timing or suppliers. But the question of what happens if your supplier's terms change is worth asking in advance:

  1. Can you export your data and settings? Who does it: the person in charge of the system. How to check: try an export on a test copy and make sure the format can be read without the service.
  2. What does the contract say about changes of terms and termination of the service? Who does it: a lawyer or a manager. How to check: find the clauses on price changes and on notification.
  3. Is there a fallback in case the price goes up? Who does it: the person responsible for the budget. How to check: name a second service and estimate the cost of switching. If there is no suitable second service or export is impossible, record this as a risk for management and decide how long without the service you are prepared to live through and what you would do by hand.

These measures are not a forecast and not a guarantee: they help if the supplier's terms change for any reason, but they do not protect against all failures.

Summary

According to Startup Fortune, citing the FT and Bloomberg, several large banks are looking for ways to pass part of the risk on loans for AI data centres to others, and a group of banks has been trying for more than six months to distribute $38 billion of debt for an Oracle project. The figures come from a retelling and need checking against the primary sources. For a company that depends on an AI supplier, it is useful to check in advance the data export, the contract terms and a fallback option.

See also another piece in this section: Nvidia is in talks to buy Reflection AI.

I work on AI agents and automation. If you would like to look at my projects or discuss your own task, visit my portfolio.

Sources