The 'free money' phase of AI bond financing has ended, as evidenced by recent lackluster demand for BlackRock's $12.5 billion bond offering. This bond, intended to finance a Meta Platforms data center in Texas, received only $20 billion in orders, significantly lower than the average four times oversubscription seen in previous AI high-grade jumbo bond sales this year. The offering yielded 7.53%, one of the highest for a blue-chip data center debt offering since the AI borrowing surge began last year, and its final spread of 2.875 percentage points over US government debt was the widest for an A-rated or higher bond in three years.
Investor concerns are mounting over the substantial debt being accumulated by hyperscalers like Meta, Oracle, Alphabet, and SpaceX to fund AI infrastructure. Companies have raised over $570 billion in AI-related debt globally since 2025, pushing borrowing costs higher and causing credit-default swaps on these companies to jump. Losses are already accumulating for some investors, with those who bought SpaceX's $25 billion bond in June now facing $1.8 billion in paper losses. This shift in market sentiment suggests a growing skepticism about whether these massive investments in AI will ultimately pay off.
Analysts predict a continued rise in financing costs for these long-term AI deals. Moody's has highlighted issues with hyperscalers' accounting practices, particularly regarding off-balance-sheet liabilities from data center leases, which have swelled to $1.2 trillion, with $820 billion yet to be fully recognized. This lack of transparency and the increasing reliance on complex financing structures like joint ventures and pre-operational leases are making investors more cautious. Despite these challenges, the demand for capital expenditures in AI is not expected to slow, with Moody's projecting these expenditures to reach almost $1 trillion next year from an estimated $785 billion this year.
