My basic model of the data-center stuff is that there’s a kajillion dollars of systematic data-center credit risk, and someone has to own it. Or in more words:

1. Companies are borrowing a kajillion dollars to build data centers for artificial intelligence.
2. The loans will get paid out of cash flows from renting the data centers to AI companies.
3. Therefore, the loans will get paid back if AI companies have a kajillion dollars of revenue over the next few decades, and if they need all those chips in all those data centers to earn that revenue.^[1]
4. If AI revenue is lower than expected, if demand for chips is lower than expected, or if chips become obsolete faster than expected, that’s bad for the data-center loans. 
5. This risk is basically all systematic, not idiosyncratic. You might lend money to one particular data-center owner (a neocloud, a hyperscaler) that leases its data center to one particular AI-business tenant (an AI lab, a hyperscaler), and you might care about the underlying credit profile of the owner or the tenant, but it doesn’t matter so much as long as there’s a real data center filled with real AI chips. If AI takes off, revenue is high and chips are in demand, but your data center’s tenant falls behind and goes bankrupt, that’s fine: The owner will lease the data center to some other AI company and never miss a payment. If the data center’s owner goes bankrupt for its own idiosyncratic reasons, again, fine: These loans are largely property-level loans against the data center, and someone will buy the data center as long as data centers are in demand.^[2] (If it goes bankrupt before building the data center, that’s maybe worse.) But there is systematic risk: If the market for data centers in 2032 is softer than expected, the data-center loans aren’t getting paid back. “Ah but my loan is backed by an investment grade hyperscaler so it will get paid back even if the AI boom is a bust,” you say, but really? All the investment-grade hyperscalers are betting a lot of money on the AI boom; in a bust scenario, are they all still investment grade?
6. Therefore, investors will own a kajillion dollars of loans that either will all get paid off, if the AI boom goes as expected, or will all be in bad trouble, if not.
7. Many of those investors — not, like, Leopold Aschenbrenner, but many of them — will want some sort of insurance against that giant systematic risk. “Hmm,” they think. “My whole portfolio has this big systematic bet on the AI boom. That is risky. I’d like to buy some insurance against the risk that it will go poorly.”
8. Lots of people want lots of things, though. Who is going to sell them that insurance? 

Like, five years ago, the world did not own very much AI data-center risk, so if for some reason you wanted insurance against AI data-center risk, plenty of people would have been happy to sell it to you. For instance, you could have bought five-year credit default swaps on Oracle Corp. — one rough measure of AI data-center risk — and it would have cost you about 0.33% of principal per year. An Oracle default was an unlikely and idiosyncratic event, and you could insure against it cheaply.

Now, the world is enormously and increasingly long AI data-center risk, so lots of people want insurance against AI data-center risk, but who would want to sell it? If you call up an insurance company and ask for $1 billion of insurance against AI data-center risk, half of that insurance company’s investment portfolio is probably in AI data-center bonds. It doesn’t want more AI data-center risk. It’s gonna charge you a lot for that insurance, and you might worry a bit about whether it could pay off: In a bad scenario for AI data centers, the insurance company’s investment portfolio will be down.

Of course you can buy “insurance” in financial markets from people other than insurers; you can call a bank or a hedge fund and buy credit default swaps on AI companies. But they all have the same problem: They are all already long tons of AI data-center risk, and will charge you a lot to take on more. (Or they are not long AI data-center risk, because they have made a conscious decision not to be, and definitely won’t sell you insurance.) Now those Oracle swaps cost about 3.3% per year.

The Financial Times reports:

Investors are racing to buy credit default swaps to protect their portfolios in the event the AI boom unravels, sending trading volumes surging for an instrument that shot to prominence during the global financial crisis.

The wave of borrowing to finance trillions of dollars of AI infrastructure spending is fuelling demand for these insurance-like derivatives, which pay out if the company they are tied to defaults on its debt.

More than $9bn worth of CDS linked to bonds issued by Oracle, SpaceX, Alphabet, Amazon, Meta, Broadcom and Nvidia was traded on average each week in September, up from less than $2bn in July, according to Morgan Stanley Research. …

In a sign of investors’ increasing unease over the borrowing binge, the cost of buying CDS on many AI-linked companies has hit record highs this month.

This so-called spread has in recent days climbed not only for companies that investors judge as riskier, such as AI cloud-computing provider CoreWeave, but also for those such as Microsoft that are seen as relatively safe. ...

“Clearly when you’re in a market where the risk is very concentrated, what you want to buy is CDS,” said Jim Caron, chief investment officer of the portfolio solutions group at Morgan Stanley Investment Management, adding that these swaps were also a “cheap hedge” for investors.

Sure of course there’s demand. But what is the supply? Eventually, if there’s going to be a $10 trillion AI buildout, it seems unavoidable that we will all own some of that risk in our retirement accounts, one way or another. (By owning AI stocks in our index funds, or by owning AI loans in the private credit funds in our 401(k)s, etc.) Or, I mean, OpenAI did once suggest that the US government should be the AI credit risk insurer of last resort. At many levels that is an annoying suggestion, but I kind of get it. This is a market-wide, systematic risk; we are kind of all in it together.

I am probably exaggerating a little. AI data-center risk is not 100% of the financial markets. Maybe it’s, I don’t know, 25%.^[3] Some people can own the market exposure to AI: About 25% of their portfolio will be in AI risk. Other people can, with great effort, get to 0%. (“Stock Pickers Burned by the AI Boom Think Their Moment Has Finally Arrived,” reports the Wall Street Journal this week; that’s nice for them.) Some people will run private credit funds or whatever, and they will invest in AI data-center loans because that’s what there is, but they will want to hedge a bit, to get their AI exposure from 25% down to 10% or something. And then other people will want as much AI exposure as they can get: 100% of their portfolio, 200%, whatever. If you are financing data centers and want to lay off some risk, they’ll happily take it from you. 

Who are those people? On the Money Stuff podcast earlier this month, I talked to Jeremy Maletz, the head of prediction markets at Susquehanna International Group, who told me that he is getting a lot of inbound demand for prediction-market products that could hedge against AI data-center risk. Well, of course he is! But who is going to take the other side of that demand? I suggested AIG, but that was kind of a joke. Leopold Aschenbrenner? (Also kidding.) Last week, Bloomberg’s Preeti Singh reported that Blackstone Inc., “the world’s largest alternative investment manager, is leaning into AI every way it can,” though I think that just means it’s financing a lot of data centers, not that it’s also selling tons of CoreWeave CDS.

Judging by the experience of the past, there will eventually be more financial engineering in this space. Some people who consciously decide to take AI risk will get concentrated AI credit risk; other people might get bonds with high ratings based on optimistic assumptions about the correlations in that risk. (“Oh, it’s a diversified portfolio of data centers.”) Financial markets can move around the AI financing risk, but they can’t get rid of it. It will be interesting to see where it lands.

Insurance companies borrow long term to lend long term. This is, I have often said, a safer business model than that of banks, which borrow short term to lend long term. You don’t get that many bank runs in insurance. People do not start worrying about the balance sheet of their life insurer and, you know, die.^[4]

On the other hand, the borrow-long/lend-long business model creates some risks of its own. If nobody needs their money back for 30 years, an insurance company can get up to some mischief with it. A lot can go wrong with an insurer’s long-term investments, over a long term, and insurance regulators try to make sure — with prudential regulation and capital requirements — that insurers make safe long-term investments so that, when their policies come due, there is money to pay them.

Of course, not all of an insurance company’s money is in long-term investments. Some people will die next week. A prudent insurance company will keep a lot of its money in long-term investments designed to earn a sufficient return to pay its long-term liabilities, and some of its money parked in short-term cash equivalents to pay current needs.

In some crude abstract approximate sense, the long-term investments are where the risk is — they’re the places where the insurance company is reaching for return to pay its long-term claims — and the short-term investments are just a safe parking place for cash and not worth worrying about.

I guess? If you take that too literally, it becomes self-defeating. A risk-seeking insurance company can be nice and sedate in its long-term investments, and take wild risks in its short-term ones, because no one is looking there. You can take quite long-term risks with short-term investments, if you put your mind to it. Lend a billion dollars to a cold-fusion company for one week, so it can take your money and immediately spend it on scientist salaries and cold fusion machines and whatever. At the end of the week, perhaps the company will have discovered and commercialized cold fusion, will be raking in money, and will pay you back with interest. Probably not, though. Probably it will have no money to pay you back. That’s fine, though. You just lend it another billion dollars for another week, and it uses that billion dollars to pay back your previous loan. (That is, you “roll” your loan for a week.) You keep doing this until it discovers cold fusion, or not.

This sounds like I’m kidding, and I am kidding, but on the other hand we talked once about an amazing trade that Greensill Capital, a supply-chain finance firm, did. Greensill was in the business of making short-term loans to companies against their receivables: A coal company, for instance, would sell some coal to a steelmaker, deliver the coal to the steelmaker on credit, send its invoice to Greensill, and get paid early by Greensill (at a discount). Then, when the steelmaker eventually paid the invoice in a month or two or whatever, Greensill would get its money back.

But Greensill would also lend money to a coal company (Bluestone Resources Inc.) against invoices that didn’t exist, because it hadn’t yet sold the coal to a steelmaker, because it hadn’t yet gotten a meeting with the steelmaker. But, you know, one day. “Greensill Capital understood that most of the Account Debtors were not existing customers of Bluestone.” These were just long-term unsecured loans to Bluestone to invest in its business. But they were documented as short-term receivables loans, and so every few months they had to be rolled: Obviously the steelmaker hadn’t paid the invoice yet, because there was no invoice yet, but one day. Eventually Greensill collapsed, these short-term loans came due, and Bluestone sued, arguing that they were actually long-term loans and so it shouldn’t have to pay them back yet. Greensill was doing an arbitrage where it told its customers that it was lending short-term, but it told its borrowers that it was lending long-term. Which was trouble.

Bloomberg’s Zachary Mider and Weihua Li reported last week on Mark Walter’s insurance companies’ short-term lending:

On paper, it was supposed to be a six-month loan. In December 2018, one of sports mogul Mark Walter’s insurance companies sent $40 million to an undisclosed borrower.

But rather than collect payment when the loan matured the following June, Delaware Life Insurance Co. rolled the debt into another six-month loan – and then another, and another. As of last December, the insurer still hadn’t collected the debt, one of dozens of “short-term” financings that have lingered for years.

The rollovers are part of an accounting maneuver that helped Walter’s insurance companies sidestep regulations while directing billions to his other businesses, according to a Bloomberg analysis of the holdings and a person with knowledge of the strategy, who asked not to be identified discussing confidential information. The move helped them reduce their required risk-based capital – the minimum cushion they must hold against the chance investments sour. …

For volatile long-term investments, such as stock or unrated loans, capital rules can require insurers to set aside 30% to guard against potential losses. But for certain short-term investments, the charge can be as low as 0.39%, no matter how risky the asset. That’s more favorable than investment-grade corporate bonds issued by the likes of Procter & Gamble Co. or Walmart Inc. …

Virtually all of the short-term assets are loans to firms controlled by Walter, typically through limited-liability companies with no public profile, according to Bloomberg’s analysis of the holdings. About $2.3 billion of the total had simply been rolled from loans involving the same LLCs in the previous year. Some of it dated back almost a decade. …

The rolling strategy was used to get lower capital charges for assets that might not be able to secure an investment-grade credit rating, or any rating at all, according to the person with knowledge of the matter.

We have talked about the Mark Walter situation several times around here. My quick summary from last week is that he “owns some insurance companies, and those insurance companies would raise money by selling annuities to customers, and Walter would control how that money was invested, and he would lend some of the money to himself to buy sports teams. Insurance regulators have pretty strict rules about related-party transactions, for reasons that are obvious if you think for a minute about the phrase ‘lend some of the money to himself to buy sports teams.’” But the insurance companies did not properly disclose a lot of these related-party loans until after federal prosecutors and insurance regulators started asking questions.

If you are going to buy, like, the Los Angeles Dodgers, you can’t really do that with a six-month loan: You plan to own the Dodgers for a long time, and they probably won’t generate enough cash to pay off the purchase price in six months. But you can buy the Dodgers with a series of rolling six-month loans, if you can find someone to make those loans to you.

One way to think about private equity is that it is basically transactional. Private equity funds are in the business of buying companies from motivated sellers (short-termist public markets, elderly founders), optimizing their capital structures and management, holding them in inventory for a while, and then selling them to motivated buyers (more enthusiastic public markets, strategic buyers). Private equity buys undervalued companies at low multiples, pulls some levers to make them more attractive to buyers and then sells them at high multiples. The profits come from the exit price.

Another way to think about private equity is just that there are some companies in the world, and if you buy good ones at good prices and then improve their operations, they will generate a lot of profits and you’ll make a nice return on your investment. The profits come from profits: You own companies that generate money, and you get the money.

The first model is a financial intermediation model. Private equity, in that model, is sort of like a slow dealer: It provides liquidity to public markets by buying companies and holding them temporarily, for a few years. It sort of implicitly assumes that there’s not too much private equity. If private equity is just going to hold companies for a few years, someone else needs to hold them for the rest of the time. 

The second model is a fundamental investing model. Private equity, in that model, is just an owner of companies. A private equity fund is almost like a classic conglomerate: It buys a bunch of disparate businesses and applies similar management and financing techniques to make them more valuable. This model is not particularly capacity constrained. If 100% of the market were private equity, private equity would own 100% of the companies, and that would be fine. There’d be no one to sell the companies to, though, except I guess other private equity companies. But you wouldn’t need to sell. You’d just collect profits from owning companies.

Anyway, a basic weird fact about the world is that “private markets” and “alternative asset managers” keep getting bigger, so a model like “private equity funds buy undervalued companies and flip them back to the market at higher valuations” just breaks down a bit. You can only do so much of that. Eventually “private equity” has to just mean, like, owning companies and collecting their cash flows. 

Bloomberg’s Laura Benitez, Preeti Singh and Layan Odeh report:

Goldman Sachs Group Inc. and Evercore Inc. are sounding out investors on a fund strategy that could help private equity firms hold onto assets for significantly longer periods.

Bankers at both firms — along with their counterparts at Jefferies Financial Group Inc., Raymond James Financial Inc. and Lazard Inc. — are pitching so-called evergreen continuation vehicles and testing structures that could appeal to select investors, according to people familiar with the matter. Talks are underway with several large private equity shops, the people said.

“We’re at the very early stages of this emergence of permanent capital CV,” Sunaina Sinha Haldea, global head of private capital advisory at Raymond James, said in an interview, declining to discuss specific transactions. “These vehicles will have more unfunded capital to support the growth of the asset, and there is no pressure to exit.”

Right, eventually the pitch is just “own this profitable company forever.”

The basic model is that tech CEOs start off kooky and wild, because that is optimal for early-stage tech companies, and become sober and responsible, because that is optimal for later-stage tech companies. I once wrote:

Early-stage startup investing is about identifying visionary ideas people, and mature-public-company governance is about making sure that companies are run by responsible adults, and it is not always the case that the visionary ideas people are (or grow up to be) the responsible adults. There is a transition, between early-stage and mature companies, that might require a change of personnel or at least a change of behavior. … If you are a certain sort of tech founder you might start taking hallucinogens because your board of directors loves them, and then later have to stop taking hallucinogens because your board of directors hates them.

Shayne Coplan is the 28-year-old founder of Polymarket and, you know, that. Here’s a Bloomberg article about how he “had been using company money to pay for private jet flights, his apartment rent and personal expenses,” but now “Private Jets, Investor Pressure Push Polymarket CEO to Grow Up”:

Interviews with people familiar with Polymarket’s operations reveal a broader effort, extending beyond Coplan’s spending habits, to bring greater discipline to a business built around the founder’s youthful entrepreneurial instincts. Some of the changes investors have sought are already in place. But the effort to turn a founder-driven startup into a mature financial exchange has suffered a number of setbacks and continues to frustrate even some prominent backers.

Senior staff at Intercontinental Exchange Inc., the parent of the New York Stock Exchange and a major stakeholder in Polymarket, have privately described difficulties mentoring Coplan, according to people familiar with the conversations. Both ICE and Founders Fund have recommended that Coplan hire a chief operating officer, a step he has yet to take, some of the people said. …

In a statement sent to Bloomberg, Coplan said: “There were real growing pains, and I personally had to grow and change into what the company needed. The only solution was hiring an excellent executive team, so I made it my number one goal for 2026.”

Sure. The other point I would make, though, is that Polymarket is not just any startup. Earlier this year, I wrote about the rivalry between Polymarket and Kalshi, the two big prediction-market companies. It seems to me that, in the early days (like four years ago), they made two very different bets on the future of prediction markets in the US. Kalshi’s bet was that, by lobbying and litigation and a general aura of upstandingness, it would succeed in making prediction markets part of the US financial mainstream, and would end up running a big regulated commodities futures exchange. This seemed like a weird bet at the time, but it turned out to be completely correct.

Polymarket, meanwhile, bet that it would be a horrible slog to get prediction markets through any sort of US regulatory approval process, so if it wanted to offer useful liquid prediction markets it would have to do everything statelessly on crypto infrastructure and sort of winkingly pretend to be off-limits to US traders. This was honestly a very reasonable bet! It just turned out to be wrong, and now Polymarket is catching up on the regulated-US-infrastructure model.

But if you were an early Polymarket investor, that was the bet you were making! You were like “ehhhh we want to bet on this guy in New York who’s running an offshore crypto casino to avoid US regulation.” And now you’re like “he shouldn’t have billed his private jet travel to the company”? 

AI Debutant’s Flameout Stokes Fears the Boom Is Ending. Anthropic turns to lawyers to help AI agents strike deals. Altman Describes Three Endgames for AI, Including One Where a Single Firm ‘Melts Everyone Else’s GPUs.’ DeepSeek Unveils Open-Source Tool to Help Cut AI Operating Costs. Real-Estate Giant CBRE Gets Deeper Into Data Centers With $1 Billion Deal. Rise of crypto vaults channels billions of dollars into risky lending. Insurers face losses as riots in France become ‘like natural catastrophes.’ Marina mania: super-rich compete for superyacht slots. HSBC reviews Palantir holding in $9bn Islamic fund. FedEx truck robbed of trading cards worth over $963,000 outside PSA headquarters. Giant Crabs Complicate New Amelia Earhart Search. 

If you'd like to get Money Stuff in handy email form, right in your inbox, please subscribe at this link. Or you can subscribe to Money Stuff and other great Bloomberg newsletters here. Thanks!