Money Stuff: Put the Data Center in the Box

The modern artificial intelligence boom has created a boom in financial engineering. A lot of people who run big tech companies and AI labs think that they will collectively need to spend trillions of dollars over the next few years building data centers to train and run AI models, buying chips to put in those data centers, buying electricity to run those data centers, and building power plants to provide that electricity.

Trillions of dollars is a lot of money for anyone, but particularly for big tech companies like Meta Platforms Inc. and Alphabet Inc., which started out as venture-backed software companies. The model there is well known: You start in a garage, you promise to change the world, you raise equity financing from venture capitalists who believe in your vision, your vision works, you scale up rapidly with very low marginal costs (just serving another web page), and eventually you become giant and profitable. When you are giant and profitable, (1) your venture capital investors get very rich and (2) you have a lot of free cash to spend on weird moonshot projects but also on the costs of running your now-quite-capital-intensive business. That is, you have a lot of cash to build data centers to serve all your web pages.

But multitrillion-dollar AI capital spending breaks that model. Those companies each make tens of billions of dollars of profit a year on hundreds of billions of revenue, but not quite trillions of dollars. And the supply of venture-type financing for AI projects is limited, in part because there aren’t trillions of dollars of venture funding — “it turns out that the supply of science-fiction-suspension-of-disbelief capital is really quite large but not trillions of dollars,” I wrote the other day — but also in part because building data centers and power plants doesn’t offer venture-type returns. A data center is a big building with computers in it. If you get a 100x return on your investment building a building, something has gone wrong. Building novel cutting-edge computer models that become superintelligent is the sort of science-fiction world-changing technological project that should get a 100x return, but renting floor space to those models is not. It should get, you know, a normal return.

And so the obvious thing to do is to sell debt to build the data centers. You know, borrow a trillion dollars, build a bunch of data centers and nuclear plants and whatnot, generate $100 billion a year of revenue from them, use $70 billion of that to pay interest on the debt, etc. Normal stable returns to pay back debt.

But the big tech companies do not want to sell debt to build the data centers, in part for real risk-management reasons but also I think a little bit for aesthetic reasons. Microsoft has Aaa/AAA credit ratings, better than the US government; Alphabet is Aa2/AA+, Meta is Aa3/AA-. Their capital-light history and huge profits have left them with very little debt and very strong credit, and they do not want to undermine that by borrowing hundreds of billions of dollars, which would bring down their ratings. [1]  

Meanwhile if you start up Joe’s Data Centers and try to raise a trillion dollars to build buildings to rent out to Meta, you will have a hard time doing that, for reasons like (1) who are you? and (2) how do investors know that Meta will actually rent your buildings? 

The problem is:

1. The AI buildout should be financed by debt investors.
2. Debt investors want to buy debt of the big tech companies, because they are good credits.
3. The big tech companies do not want to sell debt, because they are good credits.

This is a very easy problem to solve! This is known technology! What you do is:

1. There’s a box, and the box will build the data centers. (Schematically the box is often a “special purpose vehicle” or a “joint venture,” though in some cases the box is an actual public company but not a huge software company.)
2. The big tech companies promise to make some payments (rental payments, etc.) to the box. These payment obligations are not “debt,” in the sense that the tech companies’ accountants decide that they do not count as debt on the companies’ balance sheets, and the credit ratings agencies decide that they do not count as debt for credit ratings purposes.
3. The box goes out and sells debt to investors. The pitch to investors is “the big tech company promised to make payments to this box, and the box will pass those payments along to you, so if you think about it the debt of the box is just as good as the debt of the tech company.” And the investors believe that.

It is a simple sort of arbitrage: If you can convince your ratings agencies and accountants that it’s not debt, and you convince the debt investors that it is debt, then you can raise cheap debt financing without it counting against your pristine credit profile.

The details are complicated but that’s why bankers and lawyers make money. Here’s a recent Bloomberg News story about a Meta data center:

Under the SPV structure, Meta is not borrowing the capital itself, the financing entity is. Meta, in turn, will be the developer, operator and tenant of the project, which is scheduled to be completed in 2029, the people said. Through its 20% stake in the project, it’ll provide almost $6 billion of construction funding, according to a ratings report.

The SPV structure helps tech companies avoid placing large amounts of debt on their balance sheets and gives Wall Street investors the option to put money against physical assets, making them investment-grade. Structured investments are becoming more in demand as insurers and other types of investors search for debt tied to assets.

And here’s The Information on the Meta deal:

It took a year and some financial acrobatics, but Meta Platforms and a small army of bankers and lawyers were able to secure funding for a $27 billion data center project without doing any immediate damage to the company’s finances. …

Investors this month snapped up the bonds used to fund Meta’s massive Louisiana data center project, Hyperion, which is a big part of the company’s AI ambitions. Bond giant Pimco and BlackRock, the world’s largest asset manager, were among the buyers of the $27.3 billion debt offering, issued by an entity called Beignet Investor. Investment firm Blue Owl is putting up $2.5 billion in equity.

To make the deal work, Meta placed the Hyperion project into a special purpose vehicle. Blue Owl will own an 80% stake in the SPV and control its board, allowing Meta to keep the debt off its balance sheet. Meta further limited the financial impact by breaking up its leases into four-year chunks so rating agencies wouldn’t consider them debt.

To get the deal done the way it wanted, Meta walked a tightrope through accounting rules, bond rating agencies and financial regulators. Its goal was to limit its long-term liabilities, a key measure of a company’s financial health, said analysts and people with knowledge of the process. Meta obtained a letter from the SEC effectively approving its proposed accounting treatment for the project, the people said. …

Meta also reached out to credit rating agencies Moody’s and Standard and Poor’s to ensure the deal wouldn’t harm its strong investment grade rating, one of the people said. Both agencies said last week the transaction shouldn’t have any immediate impact on Meta’s credit rating, though both also highlighted risks with the deal.

That is, as far as the ratings agencies are concerned, the box that is raising money to build the data center is not Meta, so its debt is not Meta’s debt and doesn’t affect Meta’s credit ratings. (Here is the S&P report reaffirming Meta’s AA- rating and explaining why S&P “will not consolidate the debt issued by the JV into Meta’s financials.”) Meta’s backing of the box is sufficiently flexible not to count as debt. On the other hand, Meta’s backing of the box is sufficiently robust that the box’s credit rating is only one notch below Meta’s. The Information adds:

The bonds for the Hyperion data center priced with a coupon of almost 6.6%, roughly a percentage point higher than Meta’s outstanding corporate bonds and in line with the average junk bond. That’s a higher yield than investors would expect given that S&P rated the Hyperion bonds A+, safely within the investment-grade spectrum.

I should say that the big tech companies did not invent this technology to build AI data centers. This sort of thing — project finance, non-consolidated joint ventures, borrowing out of boxes — has a long history in a lot of capital-intensive industries. [2]

But it does seem to be coming into a new golden age. A lot of factors are aligning. Companies want to raise trillions of dollars of debt financing to build data centers. Those companies are very ratings-sensitive and also very concerned about maximizing equity returns, so they can really benefit from a bit of structure. The companies are full of smart engineers, so they appreciate smart (financial) engineering. And they have gotten used to paying their AI engineers tons of money in a very competitive market, so they won’t blink at paying at the top of the market for financial engineering.

Another factor, though, is the rise of private credit. Private credit firms are obvious partners for this sort of thing: They have tons of money, they really want to do investment-grade financing, and they are fine with weird structure. If you go to a traditional bond investor and say “hey we’ve got a fun bond for you but to do it you will need to invest $2.5 billion in the equity of a joint venture,” they might say “that sounds cool but our mandate doesn’t really let us own equity, particularly not in a private company.” If you go to an investment bank and say that, they might say “that sounds cool but given the current regulatory capital situation we are going to have a hard time coming up with $2.5 billion to invest in the equity of a private company.” If you go to Blue Owl and say that, they will say “yes perfect that sounds lucrative, done.” The modern suppliers of go-anywhere flexible capital are the big alternative asset managers, and there is demand for that capital. Liz Hoffman writes:

You think companies are built with equity and debt? That’s cute, today’s masters of the universe will chuckle while patting your head.

What used to be called simply “investing” or “lending” has been replaced by “capital solutions” — hybrid equity, kickers, and cash flows tailored to match the returns promised to investors on the other side. Growing pots of money now resemble liquid sand, moldable into whatever shape will fit the money hole in front of it. … And the rise of insurance money in investing has created patient capital that in many cases fits those money holes better than blunter instruments.

And while the AI boom is maybe the most obvious application for these capital solutions, it is not the only one. Once you are comfortable with the basic technology — build a box, let someone else own it, incur some obligations to the box, issue debt out of the box — there are lots of ways to use it.

Coffee for instance. Bloomberg’s Paula Seligson reports:

Keurig Dr Pepper Inc. had a problem: Investors were skeptical of how much debt the company would be taking on with its proposed acquisition of JDE Peet’s NV.

To alleviate those concerns, the company chose to tap $7 billion of financing that won’t weigh on its debt levels, including a type of investment that’s become popular for private credit players to provide in recent months.

Apollo Global Management Inc. and KKR & Co. are providing the financing, according to a statement on Monday. The firms will make a $3 billion convertible preferred stock investment in the soon-to-be-separated beverage company, while the remaining $4 billion will be an investment into a newly formed joint venture that will manufacture coffee pods. Goldman Sachs Alternatives is also investing in the venture.

Private credit funds are pumping more money into newly formed JVs as a way of providing financing, a structure that doesn’t add debt to the balance sheet of the company receiving the funds. While Keurig didn’t disclose the deal’s exact structure, companies typically promise to keep using the assets going into the JV and create a set of cash flows. Keurig’s newly formed JV will house its US and Canada single-serve manufacturing assets, with Keurig keeping a controlling interest and operating control.

Essentially, a private credit firm doles out the capital in the form of equity, but then repackages it into notes that can more easily fit within their portfolios. Most of these transactions are structured to be investment-grade, which generally attracts insurer capital. 

You set up a box, you put the K-cup manufacturing in the box, a private credit firm buys the box, you promise to buy K-cups from the box, the box issues debt. To the ratings agencies, it’s not your debt, so it doesn’t affect your ratings, but to the investors, it kind of is your debt, so they’ll buy it. It’s a general-purpose technology and it is getting a lot of use.

An interesting fact about the US listed options market is that it lends everyone money at the same rate. For instance. Today I could sell a put option on Tesla Inc. stock with a strike price of $600 and an expiration date of Dec. 18, 2026, a little more than a year from now. That option trades at around $195 per share; Tesla’s stock is around $460. [3] That option obligates me to pay $600 for a share of Tesla in December 2026. If you buy that option, you are paying me $195 now and expecting me to pay you $600 in a year for a share of Tesla. (That is not all you are doing — this is mostly a bet on Tesla’s stock price — but it is part of what you are doing.) But what if I don’t? You don’t even know who I am — you bought this option on an anonymous exchange — so how do you know I’ll be good for the $600 in a year?

The answer is, roughly, “clearing.” You don’t expect me to pay you back: You are owed the $600 by a central clearinghouse (in the US, the Options Clearing Corp.), and the central clearinghouse is owed the $600 by my broker, and my broker is owed the $600 by me. You expect the clearinghouse to pay you because it is very highly rated and carefully managed and backed by its members and sort of “too big to fail.” The clearinghouse expects my broker to pay it because my broker is a responsible, regulated, well-capitalized broker and also because my broker has posted collateral with the clearinghouse, collateral that is sufficient to cover its obligations and marked to market each day.  My broker expects me to pay it because (1) it knows what’s in my brokerage account and (2) it also has collateral from me to make sure I will meet my obligations (and to post, in turn, with the clearinghouse). The system all basically works.

So my broker will demand collateral from me to make sure I pay back the $600. But what it won’t do is charge me interest on the $600. The way this trade works is:

1. Today, you pay me $195.
2. In a year, I pay you $600 and you give me a share of Tesla. (Maybe, depending on Tesla’s stock price).
3. I post mark-to-market collateral with my broker to support my $600 obligation.

But that's it; there are no other cash flows. In a sense I owe you $600, in a sense I owe my broker $600, but I never pay interest to either of you. It’s not free credit, mind you: The price of the option implicitly accounts for the fact that I get money today and only have to pay it back in a year; there is an interest rate built into the price of the option. [4] But, because the interest rate is built into the price of the option, it is the same for anyone who buys or sells the option: If I sell that option I get $195, if Susquehanna sells the option it gets $195, so I am paying the same implicit interest rate as Susquehanna. Maybe I am posting more collateral to my broker than they are posting to theirs, but we pay the same rate.

Selling an in-the-money put option is only approximately “borrowing money,” but you can sharpen the approximation. For instance:

1. I sell a $600 strike put option for $195 and buy a $600 strike call option for $77: No matter what happens to Tesla’s stock in a year, I will be buying it for $600. [5]  And I get paid $118 today, because $600 is higher than Tesla’s stock price today: I am agreeing to overpay for it in a year, so I am getting paid now.
2. I buy a $300 strike put for $30 and sell a $300 strike call for $201: No matter what happens to Tesla’s stock in a year, I will be selling it for $300. And so I get paid $171 today, because $300 is lower than Tesla’s stock price today: I am agreeing to sell it too cheap in a year, so I am getting paid now.

The sum of these transactions is that (1) I get about $289 today (the $118 from Trade 1 plus the $171 from Trade 2), (2) next year I will pay back $300 (the $600 I pay in Trade 1 minus the $300 I get in Trade 2) and (3) nothing else. I won’t buy or sell any Tesla stock, because the trades cancel out. Tesla is not actually involved here in any way. I’m just getting $289 today and paying back $300 in a year. It’s just a loan. I’m paying about $11 for the loan, or about 3.8% of the $291 that I am borrowing.

This is called a “box spread,” and the thing to notice is that anyone can do it. Big options trading firms can do it, and I can do it, and the options trade on an exchange, and we can all get roughly the same price. We can all borrow at about 3.8%. I am probably a worse credit risk than a big options trading firm, but that doesn’t matter. If the market tried to charge me a higher interest rate — if this box spread sold for $270, say — then the big options trading firms would rush in to buy these options because they would be underpriced.

The interest rate is embedded in the price of the publicly traded options, so everyone gets the same rate. [6] The rate will vary a bit based on option market conditions but it should be somewhere close to the risk-free rate, somewhere in the ballpark of the Treasury rate. (Because, again, the magic of clearing means that this is a fairly risk-free instrument, so if it paid much more than Treasuries everyone would be doing it. We have talked about an exchange-traded fund called BOXX that lends money in the box spread market, trying to earn a Treasury-like return; if the return was way higher it would do way more of that.) 

If you are a high-net-worth retail investor and you have millions of dollars of stock in a brokerage account, you can go to your broker and ask for a loan secured by your portfolio. Your broker will be thrilled to give you a loan. What will it cost? I don’t know, but your broker will decide how much it costs, in a market that is (1) competitive but (2) segmented. Your broker knows who you are and how much this lending business is worth. If you have $1 billion in your account it will probably cost less than if you have $1 million, etc.

Whereas, if you have millions of dollars of stock in a brokerage account and you go sell some box spreads, you will just sell the box spreads at the market price of box spreads. Effectively you will borrow money from the options market at the risk-free rate, secured by your portfolio of stocks. Of course you are not a risk-free borrower, and if your portfolio of stocks goes down you will get a margin call from your broker, and if you don’t put up more collateral bad things will happen. But if your portfolio is much bigger than the loan you’ll probably be fine.

This is extremely not investing or borrowing or anything else advice, but there is perhaps a market opportunity here. Bloomberg’s Lu Wang, Yiqin Shen and Vildana Hajric report that this “strategy powers SyntheticFi, a San Francisco-based fintech” founded by a former Stripe Inc. employee named Tony Yang who borrowed $650,000 from the box spread market to buy a house in 2021 and thought “hey I could productize this”:

Box spread loans, also called synthetic borrowing, aren’t accessible to every buyer. They require sizable portfolios to back them. But for those with the assets, they offer speed, flexibility, and often a lower cost than traditional bank credit — plus potential tax advantages.

Once a tool for hedge funds and family offices, box-spread loans now sit alongside direct indexing, custom portfolios, and options overlays — all pitched as tax-efficient ways to gain financial control. For affluent investors, they’re a means to stay invested, defer taxes, and unlock liquidity without touching traditional lenders. ...

Synthetic borrow could disrupt the lending environment, according to Sam Gaeta, founder of Defined Financial Planning.

“Banks often rely on other value-added services, such as securities-backed lending, to win wealth management business,” he said. “If we can offer a solution that allows clients to access liquidity using the margin power of their investment portfolio at an implied interest rate similar to the risk-free rate, the banks lose some of that leverage in their value proposition to clients.”

Right, if you go to your bank and ask to borrow against your portfolio, they will charge you a rate that makes sense for your relationship. If you go to the options market and ask to borrow against your portfolio, it will charge you the risk-free rate.

I’m sorry I know it’s dumb but this stuff remains funny to me:

Barclays agreed to buy Best Egg, betting on consumer finance in the U.S. as part of CEO C.S. Venkatakrishnan’s bid to boost the British bank’s stock.

Barclays is paying $800 million for Best Egg, which was founded in 2013 and arranges personal loans for customers through its online platform. Delaware-based Best Egg sells the loans it makes on to asset managers, earning fees for facilitating and servicing the debts.

Barclays, which has sizable investment-banking and credit-card businesses in the U.S., aims to use the flow of loans from Best Egg to deepen relations with asset-management clients. ...

Best Egg services about $11 billion in personal loans and is expected to facilitate more than $7 billion this year. Barclays said it plans to keep a small portion of Best Egg’s new loans on its balance sheet, giving it skin in the game.

The U.S. fintech firm mostly arranges unsecured loans of up to $50,000 for customers with high FICO scores. Barclays intends to wind down Best Egg’s smaller vehicle-equity loan business, said a person familiar with its plans.

Like. You walk into the bank and you say “I would like to borrow up to $50,000 for purposes.” The banker is like “ah yes, you want a loan! Lucky for you! That is a business we just got into five minutes ago! Through our new subsidiary Best Egg! Here let me help you pull up the website and fill out the online form. If the algorithm likes your credit, we will approve the loan and get it funded by some asset managers, though of course we will keep a small portion on our balance sheet to align incentives.” And you are like “wait you are a bank don’t you just … have money … and … make loans” and the banker is like “oh you dear sweet summer child, that is not how banking works at all.” It isn’t! Now! 

I walk around New York City with some frequency, but apparently not in the right neighborhoods or wearing the right clothes, because I have never been approached by a TikToker, Instagrammer or YouTuber asking me to reveal my salary or give a tour of my apartment or name eight countries that start with A. But I also look at Instagram with some frequency, and from Instagram it definitely seems like in certain neighborhoods you can’t go five feet without being accosted for a person-on-the-street interview. 

“How old were you when you became a millionaire,” an Instagrammer called @theschoolofhardknockz asks people, for instance, and honestly that is a great opener. He seems to be doing well with it: He has more than 7 million Instagram followers, and the top three interviews on his page right now are with Tom Brady, Tom Cruise and Shaquille O’Neal. The fourth interview is from Oct. 26, and it’s with a tech executive who is not broadly famous and in fact is not named in the interview at all. But that opener still works. “Excuse me sir, how old were you when you became a millionaire,” asks @theschoolofhardknockz, and the tech executive says “ohoho this guy” and pats him on the shoulder. And answers the question! (He became a millionaire four years ago.) And keeps answering some follow-up questions, like “in what line of business” (technology) and “what do you do in tech” (“CRO for Snowflake”), even as his handler tries to pull him away. Hmm maybe … maybe don’t answer that last one? Don’t get your employer’s name into your Instagram person-on-the-street interview? I don’t have any great reason why not, just I personally would not do that. But it’s fine, it’s not like you’re giving away material nonpublic information to a person-on-the-street Instagram interviewer, right? Right? 

Theschoolofhardknockz: What did you guys do in revenue last year?

Snowflake Chief Revenue Officer: We’re going to exit this year at probably just over about $4.5 billion.

Theschoolofhardknockz: You’re going to exit this year at $4.5 billion? Are you serious?

Snowflake CRO: And we’re getting to $10 billion in a couple of years. Keep track of us.

Is that the sort of guidance that one is supposed to share in Instagram person-on-the-street interviews? No, turns out to be the answer, because Snowflake filed an 8-K the next day:

On October 26, 2025, the Instagram account named “theschoolofhardknockz” posted an interview on Instagram and related platforms with an executive officer of Snowflake Inc. (the “Company”) in which the officer made certain statements regarding the Company’s future results. Under the Company’s Corporate Disclosure Policy, this officer is not a designated spokesperson authorized to disclose financial information on behalf of the Company. As a result, investors should not rely upon such statements. The Company reaffirms the revenue guidance for Q3 and full-year FY26, originally issued on August 27, 2025, as part of the announcement of its financial results for the fiscal quarter ended July 31, 2025. The Company notes that it will release its Q3 FY26 financial results in accordance with its standard practices. The Company’s guidance philosophy remains unchanged.

Here is the full-year revenue guidance that Snowflake released in August (and reaffirmed this week), which has revenue of just under $4.4 billion (not just over $4.5 billion). I mean ... the company’s guidance philosophy has changed slightly, hasn’t it? Not the official philosophy, sure, but if the executives are going around doing Instagram interviews they might not be able to help themselves.

The stock was up about 3% on Monday and oh man if @theschoolofhardknockz traded before posting that interview, oh man. [7]

Fed Cuts Rates Quarter Point, Sets End to Balance-Sheet Runoff. Nvidia Becomes First $5 Trillion Firm as AI Rally Picks Up Steam. Morgan Stanley Buys EquityZen in First Deal for CEO Ted Pick. Cable cowboy’ John Malone to step down from media and telecoms empire. Citi probe did not interview women who complained about executive’s conduct. Tens of Thousands of White-Collar Jobs Are Disappearing as AI Starts to Bite. Business Insider asks: Is Stone Street cool?

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!