Are AI credit cracks a warning, or a ‘buy’ signal?

“Hyperion” and “Beignet” are two words many investors may not be familiar ​with. But they could soon become symbolic of the excesses of the borrowing and ‌spending binge driving the record-breaking U.S. artificial intelligence build-out.

“Hyperion” is the name of Meta Platforms’ (META-Q) AI data centre facility under construction in Richland Parish, Louisiana. It is the Facebook owner’s largest data centre project with 5 gigawatts of compute capacity, and has absorbed more than US$50-billion in investment.

Bonds sold last year to fund the project were issued through a ⁠joint venture ​in a vehicle called “Beignet Investor,” named after the famous New Orleans pastry.

As Meta’s stake is a minority one at 20 per cent, the project’s liabilities are off-balance sheet, just like some US$3-trillion of other financing and leasing structures that Nvidia (NVDA-Q), Broadcom (AVGO-Q) and hyperscalers such as Amazon (AMZN-Q) and Alphabet (GOOGL-Q) are keeping off their balance sheets, according to Morgan Stanley.

The Beignet bonds – US$27-billion of 6.581-per-cent senior secured notes maturing in 2049 – were given an A+ credit rating at issuance by S&P Global. ​However, the ratings agency also warned that “substantial credit risk” could pass to Meta during both the facility’s ‌construction and operation phases.

That risk may be materializing, as the price of these bonds is now sliding fast. On Monday, they traded at 91 US cents on the dollar, pushing the yield as high as 7.55 per cent. That’s around 230 basis points above Treasuries, a wider spread than the 185 bps at launch, but narrower than the record-wide 255 bps in July. So, no cause for panic, but it is one of many signs that cracks may be beginning to appear in parts of the AI financing edifice.

Credit default ‌swap (CDS) rates, the ​cost of insuring against default, for firms in ‌the AI ecosystem also appear to be ringing alarm bells. Meta’s on Monday climbed above July’s peak and is close to 100 basis points – and most ​of Big Tech’s CDS are also now at record levels.

On top of this, Oracle only ⁠a few days ago declared a “force majeure” notice citing potential delays in securing power for an AI data center in New Mexico, ⁠which could now be delayed by up to a year. Coincidentally, one of the other backers in Project Jupiter is again Blue Owl.

Of course, Meta and its peers are highly unlikely to ​default or go bust given the strength of their underlying businesses. This begs the question: is the current selloff in these companies’ debt logical given the scale of borrowing relative to expected revenue, or are these bonds a “screaming buy”?

Where you fall on that question – as ever with the AI story – boils down to whether you believe the returns will eventually outstrip the outlay, and how long you are prepared to wait.

The enormous investments in the AI buildout are well-documented by now. Estimates vary, but around US$1-trillion in AI ⁠capital expenditure (capex) is expected this year and around US$1.2-trillion set for next year, with economists at Oxford Economics tallying up cumulative projected AI investment of around US$3.8-trillion from 2024 to 2028. A significant portion of this is now being debt-financed, as tech firms have burned through their large cash piles.

Given these eye-popping outlays, analysts at Goldman Sachs estimate that the hyperscalers will need to generate annual AI revenues of roughly US$300-billion in the next few years to break even. Assuming hyperscalers are targeting a return on investment of 15-20 per cent, they would need to make US$570-$800-billion of additional profit, Oxford Economics reckons.

“This is hard to square with ⁠our estimates of the additional GDP that AI will create,” the firm’s economists wrote last week. They ​estimate AI will boost U.S. GDP by US$850-billion by 2032. So, at a 15-per-cent return on investment, the hyperscalers – just a handful of firms – would need to ⁠capture around two-thirds of this additional GDP, or almost all of it if we assume a 20-per-cent rate of return.

As Oxford Economics noted, “This seems very unlikely.”

Meanwhile, AI borrowing and investing continues. Hyperscalers have issued around US$250-billion ‌of debt this year, and that total is expected to rise sharply next year, even as the cost of doing so also keeps increasing. The Federal Reserve has ​started to hike its policy rate, and Treasury yields are currently the highest since before the global financial crisis. The interest burden on hyperscalers’ borrowings – both on- and off-balance sheet – is getting heavier.

How heavy is too heavy? The fate of Meta’s “deep-fried” debt deal could offer some clues.