Meta, OpenAI expand AI investments…Funding structures grow more complex [Growing AI bubble debate]

Meta, OpenAI, xAI…introduce Frankenstein finance for AI investment

Excessive borrowing and bond issuance could become a future burden

U.S. Wall Street is unveiling unprecedented financing methods amid a boom in investment in artificial intelligence (AI) infrastructure. As giant tech companies such as Meta, OpenAI and xAI pour tens of billions of dollars into building AI data centers, investment banks and private equity firms are designing "combined financial structures" that aim to both spread risk and maximize returns. However, there are concerns that complex financial structures, excessive borrowing and bond issuance could become burdens on AI-related companies and financial markets if the AI boom ends.

"Frankenstein finance" combining leases and guarantees

The Wall Street Journal (WSJ) reported on the 11th (local time) that Meta, OpenAI, xAI and others are enduring complex financing structures to fund AI investments.

The most notable case is Meta's giant data center project 'Hyperion' under construction in Louisiana. The project uses a combined finance structure that blends private equity (PE), project financing and corporate bonds, which the market calls "Frankenstein finance."

With debt already surging due to increased AI investments, Meta bypassed direct borrowing by creating a joint venture, 'Binye Investor,' to raise funds indirectly.

Blue Owl Capital invested about $3 billion to secure 80% of the private equity stake, while Meta retained a 20% stake with its already invested $1.3 billion.

The joint venture then issued $27 billion of bonds maturing in 2049, of which PIMCO purchased $18 billion. The bond yield was 6.58% annually, more than about 1% point higher than Meta's regular corporate bonds.

The key is the lease structure. Meta uses the data center as the "lessee" and pays rent, and that rent is used to pay bond interest, principal and investor dividends. However, Meta holds an option to terminate the contract every four years.

In exchange, an early termination would trigger a "guarantee clause" requiring Meta to fully compensate investors and bondholders for losses. In other words, it looks like a lessee on the books, but in substance Meta indirectly guarantees the debt.

Blue Owl described this structure as "an investment model that combines bond-like stable returns (fixed-income risk) with equity-like high profit potential." Blue Owl expects interest income based on Meta's stable rent while also anticipating equity-type gains if the data center's value rises.

Rent as the source for loan repayment

The second case is the 'Stargate' data center project jointly pursued by ChatGPT developer OpenAI, Oracle and SoftBank led by Masayoshi Son. The financing deal name for this project is called 'Jacquard.' In other words, Stargate is the data center project name, and Jacquard is the name of the financing transaction for its construction.

Vantage Data Centers, a developer and operator specializing in data centers, is building data centers worth $38 billion each in Texas and Wisconsin, and Oracle has signed 15-year long-term lease agreements. The end user is OpenAI.

However, OpenAI lacks the ability to borrow directly, and Oracle’s credit rating is relatively low among big tech firms.

Thus, JPMorgan Chase and Mitsubishi UFJ (MUFG) led the syndicate, and a banking group provided project finance loans to Vantage. Oracle pays rent to Vantage, and Vantage uses that cash to repay the bank syndicate's loans. In other words, cash flows "Oracle's rent → Vantage → the banking syndicate (creditors)."

More than 30 banks participated in this deal, and because the scale was so large some banks resold loan portions to investors to spread the risk. The interest rate is about 6.4% annually, roughly 2% points higher than Oracle's similar-maturity corporate bonds. JPMorgan received only a BBB rating from the relatively small evaluator 'Kroll' for this loan, making it difficult to include the loan in collateralized loan obligations (CLOs).

Private credit even covers chip purchases

Elon Musk's xAI applies another complex finance structure for its second giant data center 'Colossus 2.'

Musk is mobilizing cash from other companies such as SpaceX to try to surpass OpenAI, but to buy 300,000 NVIDIA chips he needs about $18 billion. As a startup, xAI cannot raise that amount all at once.

Musk enlisted the cooperation of private equity Valor Equity Partners and private credit manager Apollo Global Management. The two institutions established a special purpose vehicle (SPV), 'Valo Compute Infrastructure,' which purchases the chips on behalf of xAI, and xAI leases the chips for a set period.

Legal ownership of the chips belongs to the investor-side entity, and xAI pays rent and uses them to operate the data center. That rent becomes the source for investors to recover loan principal and interest.

In short, xAI rents the chips rather than owning them, and a chip-backed finance structure became central to the funding.

xAI chose this "chip-lease financing" for three reasons: securing cash liquidity, avoiding asset risk, and accounting benefits. If investors buy the chips and xAI leases them, initial cash burdens are greatly reduced. GPUs advance quickly, so they may become outdated after a few years.

If chip prices fall, the loss is borne by the investor who owns the chips, while xAI only pays rent, transferring risk externally. If xAI bought the chips directly, it would be recorded as large debt, but leasing treats it as operating expense in accounting and lowers the debt ratio on financial statements.

xAI reduces cash burden and risk while investors bet on the scarcity of chip assets and growth of the AI market.

However, some experts point out that "circular investment structures between NVIDIA and its customers are inflating a bubble in the AI asset market."

"If the AI bubble bursts, these will be the first to shake"

As massive financing deals for AI infrastructure construction multiply, Wall Street is filled with both strong excitement and unease. Dan Ivascyn, chief investment officer (CIO) at PIMCO, warned the WSJ: "We have not experienced a prolonged recession for a long time, and in such an environment complexity and complacency grow together," and "the volume of debt deals pouring into the market today is much larger than during past credit cycles."

New York=Shin-young Park correspondent nyusos@hankyung.com