A tidal wave of AI debt is threatening to flood the bond markets. Investors might find shelter in some surprising places.
Recent earnings reports by the so-called hyperscalers have driven expectations even higher for their collective AI spending plans. And with projections of further strained cash flows, they may need to raise a lot more debt to pay for that.
Goldman Sachs credit strategists estimate that approximately $400 billion in highly rated investment-grade bonds could be issued directly by hyperscalers globally in 2027. That is on top of the roughly $250 billion anticipated for this year.
Bond prices are often quite sensitive to supply. So these expectations have helped push down the prices, and push up the yields, of even the biggest and most creditworthy companies' bonds.
A group of investment-grade hyperscalers tracked by Barclays credit strategists includes Alphabet, Amazon.com, Meta Platforms, Microsoft, Oracle and SpaceX. The spread offered by these companies' bonds over benchmark Treasurys has widened by almost a quarter percentage point over the past month through July 30, according to tracking by Barclays.
For bond investors, one temptation might be to stick with the issuers with the highest credit ratings and to chase higher yields by focusing on their longest-term bonds. But doing the opposite might prove to be the better bet.
Already, investors can find bonds from Google-parent Alphabet that don't mature until 2075, or Amazon bonds that mature in 2065. Both of these are yielding around 6.5%, according to FactSet. Meta bonds that mature in 2065 are yielding over 7%.
Given the current pace of technological change, what the future holds for these companies so long from now is about as unknowable as anything in finance. But even beyond betting on a company's financial health decades hence, longer-term bonds have other sneaky risks. Most notably, consider what is now happening with the long end of the interest-rate curve.
Longer-term yields soared after last week's Federal Reserve meeting ended without a rate increase, or any guidance as to when one might be coming. That injected concern about inflation and the Fed's commitment to combating it into longer bonds. If investors demand much more yield because they expect future Fed rate increases or more inflation, that will have an outsize price impact on longer-duration bonds.
"You have to compare how much you get paid for what risk," says Dominique Toublan, head of U.S. credit strategy at Barclays. "The risk you have to interest rates is much, much higher on the hyperscaler long-end side."
High yields on Treasurys could also damp demand for corporate bonds that offer only a modest bump in yield over government benchmark bonds. The ICE BofA index of U.S. high-grade corporate bonds with a weighted average life of over 20 years pays about a 1 percentage point spread, according to FactSet. The cushion was twice that as recently as 2022.
Going for shorter-term, top-rated hyperscaler bonds could reduce those interest-rate risks. But it also often means getting an even smaller additional yield over Treasurys.
There is a part of the market where investors will get additional yield spread. That is in bonds that have been sold by companies to construct data centers for hyperscalers. These include bonds issued by joint ventures or firms backed by big real-estate investors, such as Blackstone and Related Cos., with tenants-to-be such as Microsoft, Alphabet and Meta.
Having hyperscalers with top-notch credit ratings as future tenants with a strong ability to pay should be a benefit. As construction risk recedes, these bonds could "look increasingly like secured, lease-backed extensions of hyperscaler credit rather than purely bespoke project financings," Barclays credit analysts wrote in a recent note. The strategists called data-center bonds the potential "sweet spot" of hyperscaler and artificial-intelligence funding.
Investors can get pickups of roughly half a percentage point to close to two points of additional yield in investment-grade data-center bonds over the bonds of their underlying hyperscaler tenants, according to Morgan Stanley strategists' tracking of a group of such bonds.
Among the risks of a data-center financing bond is that the facility doesn't get built on time or overruns its expected cost. It can be difficult to secure power, for one thing. There is also growing political opposition in some jurisdictions. The terms of the relationship with a tenant have to be carefully scrutinized, too, and there is the future risk that the tenant doesn't renew a lease.
But investors are, for now, getting additional yield spread as compensation for those risks. Some data-center bonds are also lower-rated, and with shorter-term maturities. With shorter terms, investors aren't being asked to shoulder longer-term worries about the interest rates, the U.S. fiscal picture, the Fed's credibility or even the long-term role of AI in the economy.
"With many data-center bonds, your primary risk is whether they get built or not. It's a defined risk, with an endpoint: Once the asset is stabilized, it starts to generate cash flow," Vishwas Patkar, Morgan Stanley's head of U.S. credit strategy, said in an email. "That's different from the broader, more open-ended risk that the whole AI model doesn't work out as expected."
Project bonds may also see relative price jumps as construction moves along, or potentially upgrades, too. Hyperscalers already rated toward the top of the credit spectrum might not be as likely to be upgraded. Microsoft is already rated triple-A, the highest rating. That is higher than the U.S. government.
The far-out future of AI remains hazy, and it is increasingly intertwined with huge questions about interest rates and inflation. Investors might be better paid to take the risk that at least the stuff gets built.
Comments