In August 2025, someone registered seven companies in Delaware, all with the word Beignet in their names.
Beignet, a New Orleans street food staple, is a deep-fried pastry covered in powdered sugar, guaranteed to leave a mess on your clothes when you pick it up.
No matter how many pairs of glasses you put on, you would never see how this dessert is related to AI.
A month after the emergence of Beignet companies, Meta built a data center in Louisiana named Hyperion, occupying an area equivalent to four Central Parks in New York City.
To construct this massive data center, Meta borrowed a total of $27.3 billion.
However, if you carefully examine Meta's financial report, you will find that all records related to this project on its balance sheet show only a $2.37 billion investment.
The remaining over $25 billion in debt has vanished.
This is not an isolated case.
On July 22, Nikkei Asia published a report stating that US tech giants have hidden up to $16.5 trillion in debt in invisible places, an amount even surpassing their $13.5 trillion total debt on the balance sheet.
We reviewed the filings of these five companies submitted to the SEC and found that the reality is even more sensational than that report.
On July 23, the day after the report was released, Google's parent company Alphabet submitted its quarterly filing. Its procurement commitments jumped from $332.4 billion three months ago to $811 billion. A year ago, this number was $62.1 billion.
Thirteen times in a year. This adjustment raised the Nikkei's reported $16.5 trillion to $21.3 trillion.
Over the past year, the debts created by Microsoft, Google, Amazon, Meta, and Oracle in data centers surged from $710.8 billion a year ago to $15.5 trillion. If we also consider GPU and other hardware procurement and construction contracts, it increased from $1.02 trillion to $2.86 trillion. Nearly doubled in a year.
However, only a quarter of this final number truly sits on their balance sheets. $21.3 trillion of "data center debt" has disappeared from the balance sheets of these tech giants.
Where Did All Those Debts Go?
The Bank for International Settlements had long noticed this. In a quarterly report in March 2026, it gave a name to this practice: Shadow Borrowing. The report stated that these arrangements, while economically akin to debt, mostly stayed off the company's balance sheet.
In the annual report three months later, the Bank for International Settlements once again uncommonly listed the AI Bubble and Circular Financing as primary risks to the global financial system, alongside sovereign debt.
A scan through the various filings of five companies reveals at least five different methods: SPVs, credit-enhanced derivatives, financing leases, residual guarantees, and unleased lease agreements.
And those who helped create these shadow debts are turning it into an entirely new business.
Tech Giants collectively "Return to Poverty"
For the past twenty years, tech giants have been the most comfortable type of company in the U.S. capital markets. They would make money, have cash on hand, and buy back their own stocks.
In the fourth quarter of 2021, Microsoft, Google, Amazon, Meta, and Oracle collectively repurchased $480 billion, with Meta alone spending $200 billion. With so much money, shareholders need not worry about them running out of cash.
However, in the first quarter of 2026, the total buyback amount for the five companies suddenly dropped to $46 billion.
Over the past two years, capital expenditures by tech companies in AI-related areas have more than doubled, but the growth in operating cash flow is less than sixty percent. Based on this growth rate, by mid-2027, tech giants will collectively "return to poverty" and revert back to the era of losses.
Morgan Stanley has calculated that by 2028, tech companies will need to spend around $2.9 trillion on AI, but they can generate about $1.4 trillion from their own operations. The remaining $1.5 trillion must be sourced from outside their cash flow.
So, they began to borrow. From 2020 to 2023, the five companies issued an average of about $31.3 billion in debt each year. By July 2026, this number had reached $189.7 billion, six times the previous average level.
A massive amount of corporate debt has started to strain the market. Over the past nine months, the oversubscription ratio of Amazon bonds has been continuously decreasing. In November 2025, it was 5.3 times oversubscribed, but by July 2026, it was only 1.6 times.
The issuance price has also become more expensive. In the entire investment-grade bond market this year, on average, a new bond can be sold with an additional 4 basis points. However, Amazon's July issuance required an additional 18 to 21 basis points to sell.
Google and Oracle took it even further by directly raising money through stock issuance, raising nearly $80 billion and then introducing a combined $60 billion ATM issuance plan.
Financial media collectively complained that tech giants have violated the "implicit contract" with investors. Previously, when the market bought shares of these companies, it assumed they had net cash, low debt, and continuous buybacks. Now, they are heavily issuing debt and have to halt their buyback programs.
However, the real trouble is reflected in the balance sheets of these giants.
The more debt on the balance sheet, the more cautious rating agencies become, and the fewer people are willing to buy the debt. The first of the big five to hit this wall was Oracle. Its capital expenditure jumped from $21.2 billion to $55.7 billion within a year. In July 2026, S&P downgraded it from BBB to BBB−, just one notch above junk status. If it is downgraded another level, global insurance companies and pension funds will be required by regulations to sell Oracle's bonds.
Thus, these giants not only need more money, but they need money that is discrete, long-lasting, and less onerous. And this kind of money is simply not found in the open market.
After SaaS "Breakdown," Wall Street Seeks New Business
Just as tech companies are worried about money, on the other side of Wall Street, someone is looking for a way out.
In the first half of 2026, a fund under Blue Owl received redemption requests nearing 40% for two consecutive quarters, but only managed to pay out just over 10%. The company's stock price plummeted from $24 to $9.
Blue Owl is one of the world's largest private credit institutions, managing over $310 billion in assets. The fund facing redemptions specializes in the software sector, with software loans accounting for over 60%.
Ironically, in 2026, Wall Street's least desired asset is software loans.
Starting from the peak in October 2025, software stocks have dropped by almost 40%. The market's explanation is straightforward: AI will kill software. In the past, investors were willing to give software companies high valuations because customers renewed annually, and revenue seemed to continue growing along the contracts. Now, whether customers will renew has suddenly become a question.
However, in reality, the software company's fundamentals are not that bad. The revenue growth of Microsoft 365 subscription business has increased from 15% to 19%, and the revenue growth of software companies such as ServiceNow, Salesforce, and Snowflake has accelerated for five consecutive quarters. Gartner has also raised its forecast for global software spending.
But in the financial world, confidence is often more important than fundamentals.
In a shareholder letter, Blue Owl admitted that the market's concerns about AI impacting software companies have significantly influenced how investors view software credit exposure.
Wall Street desperately needs a new story to get investors back on board. And that story is the data center.
Software lending bets on whether customers will renew next year. Data center bets on whether AI companies will need computing power. The former question is becoming increasingly difficult to answer, while the latter question seems almost unnecessary to answer. The stronger AI gets, the more valuable the data center becomes.
Now, data center lending is the hottest business on Wall Street. In December 2025, Blue Owl rejected Oracle's Michigan data center project on the grounds of not meeting underwriting standards. However, Pacific Investment Management quickly snatched up the deal at a higher price. Such bidding wars occur almost every month on Wall Street.
On one side, there are cash-strapped tech giants, and on the other side, asset management institutions that have nowhere to invest their money. So, they strike a deal.
Their first major project is Beignet, the dessert mentioned at the beginning.
How did Meta hide $28 billion in debt in a dessert?
The Hyperion hyperscale data center located in Louisiana, built with Meta's involvement, was first registered under Laidley LLC. The park is operated by it, and it also signed a fifteen-year power supply contract with the local power company.
Laidley belongs to Project Beignet Holdings, a joint venture company. The ownership of the data center is here.
The majority shareholder of the joint venture is called Beignet Investor. $27.3 billion in bonds were issued from its hands.
The debt is not placed on the company that owns the data center, but on its shareholders.
Further up, there is the Beignet Pledgor. In English, pledgor is the pledger. It wholly owns Beignet Investor, which in turn pledges all of Beignet Investor's shares to the trustee.
This allows the creditors to have a very straightforward collateral. In case of a problem, the trustee does not need to first assess a data center in Louisiana, sell a data center, or wait through a long litigation process. They can simply enforce the shares per the contract. The campus is still running, the lease is ongoing, and the rent collector has changed.
Above Beignet Pledgor, there are four more companies, one of which is named Beignet Net Lease Aggregator. At the very top is a Net Lease Real Estate Trust called OSNL under Blue Owl, along with co-investors who contributed alongside it.
All seven companies are registered in Delaware. Local LLCs are not required to disclose members and ownership percentages.
The $27.3 billion debt did not go through a public offering. It took the 144A route, only selling to qualified institutional buyers without submitting a prospectus. To review the terms of the transaction, one must first sign a confidentiality agreement. It remains a perpetual 144A offering and will not convert to publicly registered bonds.
In the SEC's full-text search system, searching for "Beignet" yields only one descriptive hit, which is in the Blue Owl quarterly report's subsequent events footnote. By the time the annual report comes around, it and "Meta," along with the name of the county where the project is located, disappear, leaving only a single line summarizing "Net Lease Data Centers."
That is already enough to make one's head spin, but it is merely legal isolation.
SPVs are nothing new. The real estate and infrastructure industries have used them for decades, and accounting standards have long anticipated that people would stuff debt here, so they have left two thresholds. Whether an entity should be consolidated into a company's financial statements depends not only on the ownership percentage but also on who can control the most critical operating activities, who bears most of the losses, and who takes most of the gains.
Meta is Hyperion's sole tenant, providing funds, credit, and responsible for construction and property management. By this criterion, this debt should logically be consolidated into its own financial statements.
Yet it retained 20%.
Blue Owl's OSNL fund and the co-investors, through Beignet Pledgor and the other four holding companies, own 100% of Beignet Investor. Beignet Investor then holds an 80% stake in a joint venture company. Meta holds the remaining 20%.
80 and 20 are the two numbers that repeatedly appear in this story.
Within Meta's financial report, the company stated that it does not have the power to dominate activities that most affect the performance of joint ventures, so it is not the primary beneficiary and is not consolidated. Joint ventures are not included in Meta's financial statements, and the $27.3 billion debt naturally is not included either.
The account has not disappeared. It has just changed its storage location.
Lease Accounting, the "Debt Repayment Art" of Silicon Valley Giants
The fact that the debt is not on Meta's books does not mean that Meta does not have to pay.
Meta owns a leasing subsidiary called Pelican Leap. It signed a four-year lease with Laidley. Starting in 2029, Pelican Leap will pay rent to Laidley every month. The money goes from Laidley to the joint venture, then to Beignet Investor, and is then used to repay the bondholders' principal and interest.
After going around in a big circle, the rent still ultimately comes from Meta's pocket.
The $27.3 billion bond has an interest rate of 6.581%, matures in May 2049, and is amortized in full. It does not wait until 2049 to repay the principal, unlike regular corporate bonds. A little piece is paid off each period over twenty-four years.
It is more like a mortgage.
The total rent that Meta has to pay in the first four years amounts to $12.3 billion, averaging $3.08 billion per year. This figure conveniently covers the principal and interest due that year, with an additional 12% left for equity contributors.
The real drama begins with the lease term.
The bond has a term of twenty-four years. The initial lease term is only four years. Starting in 2029, there is an option to renew attached, which can be extended for up to twenty years. By 2033, Meta can theoretically choose not to renew and simply walk away.
But what about the remaining over $20 billion?
The answer lies in another part of the lease. In addition to the monthly rent, Meta has provided residual value guarantees, with a cap of around $28 billion, slightly higher than the debt itself, decreasing over time. If Meta does not renew the lease, Meta will make up the difference for the portion below this threshold of the park's value.
When putting $28 billion and $27.3 billion together, it is hard not to associate the two.
So what the bond truly relies on is not just the building, nor just the machinery inside.
It relies on Meta's credit.
This also explains the rating. S&P rated this bond as A+, while Meta itself is AA−. The rating agency did not price the bond based on a building that is yet to operate, but instead downgraded Meta's credit by one notch.
Meta is providing the money, the credit, is responsible for the operation, and is the sole tenant, yet in the financial statements, it states that it is not the primary beneficiary.
This clean balance sheet doesn't come cheap. If Meta were to issue bonds of the same tenure in the public market, the cost would be around 5.5%. Through this structure, the cost rises to 6.581%. With the same amount of money, it would pay close to $300 million in interest in a little over a year.
Those willing to spend this much money to acquire it are evidently not just buying a building.
In July 2026, a second similar project arrived, named Sopaipilla, after a type of fried pastry common in the American Southwest. The project is located in El Paso, Texas, with an initial bond issuance size of $12 billion, also at an 80/20 split. The difference is that 80% of the funding has been switched from Blue Owl to BlackRock.
Meta has given two internal codenames to the next-generation large-scale model, one is Avocado, and the other is Mango. The financial side's naming is evidently more appetizing.
Meta's structure is the most intricate, but it is not the only approach.
Microsoft did not use a shell company and did not significantly increase its book debt. In the past two years, its total debt decreased from $44.9 billion to $40.3 billion. However, during the same period, finance lease liabilities surged from $27.1 billion to $62.9 billion, more than doubling, exceeding the total debt by over $20 billion.
Indeed, the $62.9 billion is on Microsoft's balance sheet, just not under "Debt" but split into "Other current liabilities" and "Other long-term liabilities." From the most visible line, it remains muted.
The meaning of finance leasing is also straightforward. It is nominally a lease but closer to installment purchase. The lease term covers most of the asset's useful life, accounting-wise equivalent to ownership, just with installment payments. Therefore, it has to be fully recorded as a liability but does not have to be reported under debt.
Google does not use a shell either. It provides payment guarantees for data centers built by others, enabling them to borrow money. In the financial statements, Google left a crucial phrase, reserving the right to take over the underlying lease if the counterparty defaults.
These kinds of guarantees are recorded as credit derivatives. The notional size increased from $16.9 billion to $43.8 billion within six months, of which the amount truly entering the balance sheet is the guarantee's valuation at the end of the day, $815 million, less than two percent of the notional size.
Amazon seems the most straightforward. In March 2026, it issued over $50 billion in bonds to build its own offices. It also has $106.3 billion in operating leases, similar to Oracle, which have not yet hit the balance sheet.
Oracle's approach is simpler; it signs contracts. The $260 billion in operating leases are mostly data center-related, with terms of fifteen to nineteen years, and will start being recognized in the balance sheet in the 2027 fiscal year. Until then, these lease liabilities will not be on the balance sheet.
Meta uses joint ventures. Microsoft uses financing leases. Google provides credit guarantees to projects. Amazon and Oracle opt for signing leases first.
Although each is taking a different path, everyone is focused on the same line.
By roughly dividing the future payments a company has to make into three layers, you can see that line. The first layer is money already borrowed, such as bonds, notes, loans - money in, debt on the books. The second layer is obligations formed after goods, services, or assets have been received or utilized, including rents already being paid. The third layer consists of contracts already signed for services and assets not yet delivered or in use, usually disclosed in the footnotes of financial statements.
Between the second and third layers lies the boundary of the balance sheet.
Accounting doesn't ask whether you intend to pay back the money in the future; it only asks if you have received something now. If you've received money, goods, or a usable building, the account has to reflect it. If you haven't received anything yet, the debt may not need to be recognized immediately.
That's the whole story. Don't buy a building outright, sign a lease. Don't let the lease start today, wait a few years.
Leases also vary in depth. Financing leases go on the balance sheet but are buried within other liabilities, such as Microsoft's $62.9 billion in this layer. Operating leases also hit the balance sheet but only a small portion in the present value of rents, Meta's leases fall into this category. Leases not yet commenced do not hit the balance sheet at all, like Oracle's $260 billion and Amazon's $106.3 billion standing on the outside.
The act of building, over time, has morphed into leasing.
When you add up the numbers of these five companies, the money already borrowed amounts to $445.8 billion. Leases already signed but not yet recognized total $831 billion, nearly twice as much. When you factor in procurement and construction commitments, the total comes to $21.3 trillion.
This is the magnified version of that initial couple hundred billion.
5 years ago, Wall Street had already set its sights on the data center business
If the story ended here, it could easily be misconstrued as a new invention spurred by the AI frenzy.
Actually, no.
Let's rewind to 2021. Global data center M&A reached $49 billion, setting a new record at the time. In 2022, it was $48 billion, with 187 deals, of which 91% of the funding came from private equity. Out of the top twelve largest deals that year, ten were acquired by private equity.
Blackstone acquired QTS Realty Trust for around $10 billion. KKR and GIP acquired CyrusOne for $15 billion. DigitalBridge and IFM acquired Switch for $11 billion. The average deal size increased from $80 million in 2018 to $235 million in 2022.
What they were buying was not servers.
It was land. It was buildings. It was the incoming power.
In 2023, with the interest rate hike, global data center M&A shrank to $26 billion. By 2024, it surged to $73 billion, surpassing all previous records. It hit another high in 2025. From early 2024 to now, there have been 575 deals totaling $151 billion, with 84% of the funding coming from private equity.
Because the tenants are solid, and the contracts are long.
Data centers burn money, and often, data center operators cannot fund the builds themselves, so they seek partners. Yet, the ones moving in are the highest-credit cloud giants, signing leases for over a decade. For infrastructure funds, pension funds, and sovereign wealth funds seeking stable returns worldwide, this is not a tech game; it's more like a fully powered rent-collection building.
Even the equity arrangement for Hyperion has been used in the market before.
In December 2023, Blackstone and Digital Realty formed a $7 billion development joint venture. Blackstone held 80%, and Digital Realty retained 20%. In October 2024, Equinix, the Government of Singapore Investment Corporation, and Canadian pension funds established a joint venture exceeding $15 billion, with Equinix holding a 25% share.
Blackstone's 80 and Digital Realty's 20, as well as Blue Owl's 80 and Meta's 20, saw little change.
The structure was set long before AI. The assets are there. Long-term tenants are in place. Even the proportions are set. Only thing missing is who will put up the ever-increasing sums of money.
In theory, it should be the turn of the banks.
The banking sector, however, took a step back in 2023. Following the collapse of Silicon Valley Bank, US regulators served up a draft of the "Basel III Endgame" capital rule in July of the same year. For long-term, large-scale, and highly customized loans, banks will need to hold more of their own capital.
The data center precisely hits on three words. Long. Big. Non-standard.
Banks have started to deleverage, with private capital stepping in to fill the void. Behind the latter are insurance annuities and pensions, with fund durations aligning with twenty-year leases.
This baton pass is particularly crucial for private credit. Over the past decade, the biggest story has been software, with software loans growing from less than $80 billion in 2015 to over $500 billion by the end of 2025, accounting for 19% of all direct loans. AI-related loans have surged from nearly zero to over $200 billion, with their share increasing from under 1% to close to 8%, primarily occurring in recent years.
The proportion of private credit funds invested in AI-related fields has risen from 5% a decade ago to 20%. In terms of transaction volume, AI-related deals now represent 34%, compared to an average of only 17% in the previous five years. 144A private placement bonds in the data center sector, which were almost non-existent by the end of 2025, have now grown to over $40 billion. Over the next three years, AI infrastructure is expected to absorb $800 billion from private credit.
While it took a decade for software to amass its volume, AI aims to catch up in less than three years.
However, whether private capital can maintain this position remains to be seen. In March 2026, the requirements of the "Basel III Endgame" were significantly relaxed. The risk weightings for corporate loans have been reduced, and the capital charges for private fund investments have been restored to four times the level in the draft. Regulators estimate that this will release over $1 trillion in additional lending capacity for banks. JPMorgan Chase has already set aside a $50 billion direct lending facility.
The banks are gearing up to return.
The Subprime of the AI Era?
Now, the market has given a new name to these shadow loans: the "Subprime of the AI Era."
Before the 2008 financial crisis, everyone assumed that home prices would keep rising. Today, there are also two "defaults." The first is the assumption that data centers will be completed on time. The International Monetary Fund estimates that 60% of planned data centers have yet to break ground, yet the related debt has already been sold. Hyperion is expected to be completed around 2029, Oracle's $26 billion lease is set to commence from the 2027 fiscal year, and the Sopaipilla Campus aims for completion in 2028.
Secondly, once the building is completed, the machines inside should still hold sufficient value.
The maturity date for Hyperion's debt is 2049, twenty-four years from now. The graphics cards in the project are typically depreciated over 5 to 6 years on a tech company's balance sheet. Short sellers believe the true lifespan is only two to three years. In the second-hand market, the H100 can only fetch 45% of the original price by the third year.
The market has been trying to assess this unease. The cost of buying default insurance on corporate debt reflects the market's view of default risk. Oracle's CDS price exceeded the peak of the 2008 financial crisis in March 2026, setting a new record four months later.
Today's valuation is not based on a stable, rent-paying building but on a series of events yet to occur. Whether the building will be completed on time, electricity will be connected as planned, computing power will be utilized, machines will age less rapidly over the twenty-four years of debt repayment, and tenants will renew their leases...
The fund itself has also begun to leverage. In the second quarter of 2026, investors requested redemptions of $15.6 billion but only received $5.9 billion. During the same quarter, Apollo, BlackRock, Ares, and Blue Owl successively tapped the debt market for financing.
Existing investors are unable to exit, yet the fund continues to borrow new funds.
However, during the subprime mortgage crisis, the borrowers who couldn't repay their mortgages were ordinary families. Now, it is the highest-rated companies globally that are borrowing. The question is, can they uphold their credit and commitments?
Welcome to join the official BlockBeats community:
Telegram Subscription Group: https://t.me/theblockbeats
Telegram Discussion Group: https://t.me/BlockBeats_App
Official Twitter Account: https://twitter.com/BlockBeatsAsia
