AI costs are rising! Meta raises another $12 billion, borrowing costs significantly increased

As the investment frenzy in AI infrastructure by tech giants continues, the bond market is pricing this costly competition at higher rates. Meta’s latest $12 billion data center financing shows that investors are demanding a notably higher risk premium for AI-related debt, with borrowing costs having risen significantly compared to the previous transaction nine months ago.

On Friday, according to Financial Times citing sources familiar with the matter, the bonds—led by BlackRock and used for a near-1GW data center project in El Paso, Texas—had an initial yield discussion of over 7%, with the risk premium demanded by some investors about 0.4 percentage points higher than Meta’s “Hyperion” data center deal last October. The sources added that pricing discussions are still in the early stages, with the issuance expected to formally launch as soon as next Monday, and final terms may change.

The rise in borrowing costs reflects the bond market’s growing caution towards AI-related financing risks. Meanwhile, AI stocks have recently suffered large-scale sell-offs, and equity investors’ doubts over the sustainability of the sector’s boom are intensifying, resulting in resonating sentiments across both markets.

Borrowing Costs Climb as Market Reprices AI Risk

The cost pressure of this financing is significant in the bond market. One credit investor specializing in investment-grade debt commented: When you are issuing tens of billions in bonds, even a 0.1 percentage point rise in cost results in tens of millions of dollars in additional annual interest expenses, which is very impactful in the high-grade market.

For reference, Meta’s previous “Hyperion” project bonds completed a record $27 billion issuance last October, and the bonds linked to that project—issued via a special-purpose vehicle called “Beignet Investor”—have fallen to about 96 cents on the dollar as of this Thursday.

According to sources, this increase in borrowing costs directly reflects lenders’ cautious attitude towards their expanding AI exposure—a caution that is intensifying following a borrowing spree among tech giants over recent months.

Special Purpose Vehicle Structure Keeps Tech Companies’ Balance Sheets “Clean”

This financing continues the structural design of Meta’s previous deal. The bonds will be issued through a special-purpose vehicle named “Sopaipilla Investor”—named after a popular South American fried pastry, in line with the previous “Beignet Investor” that took its name from a Louisiana specialty dessert. Sopaipilla will hold 80% of the Texas project’s equity, with the remaining 20% retained by Meta.

S&P analyst Viviane Gosselin stated, this deal is “almost an exact replica of the last transaction.”

Raising debt through project entities rather than at the corporate level has become the mainstream method for tech companies in the AI arms race to secure funding while keeping their balance sheets clean. Last month, Anthropic also completed $35 billion in financing with a structure backed by GPU leases and a Broadcom guarantee.

Robust Structure, But With Specific Risk Clauses

Regarding bond structure, the Sopaipilla bonds will mature in 2048, backed by Meta’s 20-year rental payments from 2028. Meta retains a renewal option every four years, up to four times in total, and would be required to pay a substantial break fee if terminating early, providing significant protection to lenders. Meta also assumes construction risk and must cover any overruns exceeding 105% of the initial budget.

However, the structure has its limitations: the bonds are not directly secured by physical assets. S&P noted in its report that if the project suffers a major accident causing a delay of more than 18 months, Meta can terminate the lease without any penalty.

As for ratings, S&P assigns the bonds an A+, one notch below Meta’s corporate rating of AA-. Gosselin said: "From our perspective, this is a very robust structure." Fitch and KBRA have rated the transaction at AA-, on par with Meta’s corporate rating.