Shuli Ren, Bloomberg
Ambition comes at a cost. As tech companies race to build out their AI infrastructure, those doing creative financing deals are being penalized by credit markets.
Exhibit A is Meta Platforms Inc.’s massive data-center project in rural Louisiana, financed off-balance-sheet. The company formed a joint venture with private equity firm Blue Owl Capital Inc. and took a 20% minority stake. And it was Beignet Investor — a special-purpose vehicle with majority interest in the joint venture, not Meta — that borrowed $27 billion. Before closing the deal, the tech giant had reached out to Moody’s Ratings and S&P Global to make sure this structure wouldn’t hurt its investment grade.
While the bonds were privately placed, Beignet ended up selling them to Main Street funds such as Pacific Investment Management Co. and BlackRock Inc. Pimco had committed $18 billion to this deal, but demanded a higher yield and guarantees in exchange. The 25-year bond pays a 6.6% coupon, about a percentage point higher than Meta’s corporate bonds of a similar tenor.
Meta’s bonds sold off shortly after. This is because fixed-income managers such as Pimco, which can invest across public and private debt, nonetheless sees Beignet as a Meta subsidiary. After all, the tech giant has agreed to cover its joint venture’s outstanding debt should it choose not to renew the lease of the data center. Having bought billions of dollars of Beignet’s bonds, asset managers have to de-risk and sell some of their Meta holdings.
Or consider Oracle Corp., by far the most aggressive hyperscaler. Capital expenditure is expected to reach 138% of its operating cash flow next year, well exceeding runner-up Meta’s 84%. In recent days, the cost of its five-year credit default swap, essentially an insurance for corporate debt, has jumped.
The spike is probably due to hedging demand from banks and Oracle’s business partners. Dozens of lenders have been busy arranging jumbo-sized loans to finance the ambitious Stargate project, spearheaded by OpenAI Inc., SoftBank Group Corp. and Oracle to rapidly build artificial intelligence infrastructure across the US. They include a $18 billion deal to fund the construction of a campus in New Mexico, as well as a $38 billion package for sites in Wisconsin and Texas. Oracle is expected to be the tenant.
Lenders have good reasons to fret. Building an AI data center costs about $50 billion per gigawatt, with as much as 30% of the expenses attributed to land and facilities, according to Barclays. To scale up, Oracle has chosen to save upfront costs, lease the land, and concentrate its spending on chips instead.
Ratings agencies like to go by the book, but the credit markets have already moved on and made their own judgment and adjustments. At the beginning of the year, blue-chip tech companies were able to sell bonds more cheaply than their peers; this is no longer the case. Among investment-grade tech firms, Oracle’s bond spread has widened the most. Earlier this week, when Amazon.com Inc. sold $15 billion of notes, investors withdrew 40% of their orders when the final pricing failed to offer enough yield.
A company can play with shadow banking if it’s in the shadows. But the hyperscalers’ deals are so big that every tidbit has to be carefully scrutinized. Their financial acrobatics is fooling no one. [Abridged]
Courtesy Bloomberg/Shuli Ren
No Comments