The huge lurking leverage of AI hyperscalers

The writer is a bond portfolio manager at Barksdale Investment Management and editor of ‘The Credit Investor’s Handbook’

At times, creditors can feel like they are the only adults in the room. As AI-driven exuberance drives stock markets higher, bond vigilantes are pushing Treasury rates higher while credit markets are repricing data centre-related debt lower.

While equity analysts might get paid for dreamy earnings forecasts, the best creditors can hope for is repayment at par if a bond is held to maturity.

So they look at the details where, as ever, devils reside. For example, while the debt being taken on by hyperscalers to build out AI infrastructure is getting lots of attention and is even being blamed for crowding out Treasury bonds, creditors have taken this a step further in typically geeky fashion, scrutinising off-balance-sheet obligations such as leases.

Any credit analyst worth their salt puts lease obligations into their balance sheet calculations, capitalising them in financial jargon — even if the issuer does not. We don’t do this to be schoolmarms; like bond coupons, operating leases are typically non-cancellable outside of bankruptcy. Senior unsecured creditor committees for retailers in Chapter 11 routinely have a landlord representative sitting at the same table as unsecured bondholders and trade creditors — these three creditors are “pari passu”, ie equal, in their claims.

As outlined in The Credit Investor’s Handbook, the rule of thumb after the introduction of the ASC-842 accounting rule in the US is to add the balance sheet operating lease liability to debt and divide by ebitda plus rent to arrive at “lease-adjusted leverage.”

As hyperscalers move from cash cows to net borrowers, it’s easy to account for the investment-grade-rated debt issued to build data centres. The dollar amount, maturity and coupon of bonds are fixed and clear. When Meta issued $25bn in bonds earlier this year, analysts duly added $25bn to its debt burden and the new interest expense to their cash flow forecasts.

But when Meta and private capital firm Blue Owl created a venture last year to build a huge data centre in Louisiana known as Hyperion, the maths was not so easy. This and other hyperscaler lease obligations are murky, in part because most of the associated data centres are in the construction phase and thus not yet reflected in the balance sheet operating lease liability.

What we do know about them is not reassuring. For example, Meta disclosed in its most recent quarterly report some $347bn in future lease obligations on projects that have not yet commenced. This compares with a puny $84bn of on-balance sheet debt at the end of 2025.

The $347bn is not the balance sheet operating lease liability; Meta will use a discount rate to calculate the present value of the liability of each lease as it comes online. But it would take a double-digit discount rate to whittle that below $100bn.

Comprehensive lease disclosures are only required annually. So in a year like 2026, which has seen an explosion of data centre leases, creditors might find an unpleasant surprise in Meta’s next annual report. Even once it is filed, the disclosure may not give them much to go on. Unlike store leases which are fairly standard, if you’ve seen one data centre lease, you’ve seen just one data centre lease.

Recommended

Although Meta has not provided explicit capex guidance beyond 2026, sell-side estimates exceed $500bn through 2028, more than eating up its forecast operating cash flow. Whether the associated financing takes the form of lease liabilities or debt, Meta and its affiliates (future vehicles like Hyperion) will return regularly to the credit markets over the next couple of years.

All of these moving parts — debt and lease obligations, along with growth in Meta’s ebitda and operating cash flow — make it challenging to get a handle on its balance sheet trajectory. The range of outcomes for Meta’s lease-adjusted leverage by year-end 2028 is wider than typical for a double A-rated credit. It’s no wonder its bonds — which my firm does not hold a long or short position — trade more like debt of Verizon rated high-BBB. Price-earnings ratios show no such parity — Meta’s is more than twice Verizon’s.

History shows that the stock market tends to ignore (or, even, prefer) off-balance sheet obligations until an inflection point is reached. Creditors may not be sure exactly how to account for the lease obligations Meta has incurred, but their best guess will appear somewhere in their models — and, ultimately, in the spread they require to own Meta bonds. By the time equity analysts start inviting their bond counterparts to “grab a beer to pick your brain about leases”, hyperscaler bond prices — and PE ratios — may be a lot lower.