AI has taken over the stock market. The bond market is next

The latest techno-infrastructure frenzy, this time over artificial intelligence, is likewise stoking a bond-market boom (see charts 1 and 2). Since January Meta, Nvidia and Oracle have launched individual bond offerings worth $25bn apiece to fund their AI ambitions—more than any of them has ever raised in equity. So has SpaceX, which added the same sum to the $86bn that Elon Musk’s rockets-to-robots business raised from stock-market investors last month. In March Amazon, whose computing cloud hosts lots of AI workloads, sold $37bn in bonds. Alphabet issued a rare £1bn ($1.4bn) 100-year bond in Britain in February as part of a £5.5bn debt sale, plus more paper in Swiss francs, days after it raised $20bn at home.

Morgan Stanley expects $350bn-400bn in AI-related investment-grade issuance this year in America. That would be nearly a fifth of a record $2.3trn in high-quality dollar-denominated bonds that the bank forecasts firms will sell in America. Another $50bn may come from junk bonds linked to AI projects (out of total high-yield issuance of $440bn). The Bank for International Settlements, central banks’ central bank, recently warned that AI projects may not make enough money to repay the debt that financed them. If bond investing is the “negative art” of choosing what to avoid rather than what to buy, as it is sometimes described, how much avoiding must bond investors do?

In some ways, the AI bonanza makes the corporate-bond market appear a little safer. The total debts of America’s five “hyperscale” cloud giants—Alphabet, Amazon, Meta, Microsoft and Oracle—climbed by $228bn in the six months to March. That is nearly five times more than any such two-quarter increase in the past. The first four of the firms have for years churned out more profit than they know what to do with, so enjoy strong credit ratings. Microsoft’s debt is considered safer than Uncle Sam’s. Only Oracle, the smallest and least profitable of the five, receives a middling “B” grade from big credit-rating agencies (though so do the issuers of about half of American corporate bonds these days).

You might expect a debt-raising boom to raise the spreads on corporate bonds, a measure of their risk relative to safe Treasuries. But thanks to the creditworthiness of the biggest issuers, spreads remain at around 0.8 percentage points on average, close to the lowest in a quarter of a century.

Investors are also beginning to be more discerning between the safe debt of the hyperscalers and bonds funding some of the riskier projects with which the tech giants are associated. In April QTS Data Centers, a firm owned by Blackstone, a giant manager of private assets, issued $4.6bn in bonds to fund an immense data centre in Georgia, of which Microsoft will be a client. Initially the bonds’ yield was 1.1 percentage points higher than for the equivalent Microsoft debt. Since then the spread has widened to 1.6 points.

On the surface, the ballooning issuance of riskier debt looks like a bigger cause for concern. The wave of junk bonds, set off last May by a $2bn bond offering from CoreWeave, which rents out top-tier chip capacity, is swelling. In addition, CoreWeave and fellow “neoclouds” like Nebius and Iren have all issued billions of dollars in convertible bonds this year, which investors can swap for a certain number of the issuer’s shares.

The division between safer and riskier debt may be misleading, however. Historically, financial instability has tended to build up around assets that investors believe to be safe rather than those they already see as sketchy.

When railway defaults surged in the wake of the financial panic of 1873, it led to a worldwide depression. No such cataclysm looks imminent today. It is clear, though, that avoiding duds will be harder than ever. The AI revolution is proceeding at such a blistering pace that some early winners have already turned into losers, then, occasionally, back into winners—and vice versa. Divining which companies will come out on top, and be able to redeem their bonds when they mature in 10, 30 or, in Alphabet’s case, 100 years’ time, will require more negative artfulness than ever.