How AI conquered the bond markets

In 2026, the five hyperscalers (Microsoft, Amazon, Alphabet, Meta and Oracle) will spend around USD 820 billion in capex, compared with USD 750 billion in operating cash flows. This means that their investments are exceeding available cash flows and that this funding gap has to be financed through debt.

Dollar-denominated investment-grade (IG) bonds cover most of this financing need, with the rest coming from bonds in other currencies, project financing, leveraged loans, securitisation and chip-backed financing. On top of this, there is close to USD 3 trillion being held off-balance-sheet, in lease agreements on data centres under construction, purchase commitments, and guarantees to less secure stakeholders.

The borrowers are among the most solid in the market: Microsoft holds one of the rare AAA corporate credit ratings, while Alphabet, Amazon and Meta are each rated AA- or better. These four hold more than USD 500 billion in cash, their revenues are growing by around 20% per year, and their debt levels are well below S&P’s downgrade thresholds. Together, they could issue USD 400 billion in debt without troubling their ratings. However, the real limit is in the markets’ appetite for debt: the five hyperscalers (plus Nvidia) already account for 4.6% of the US IG index, and premia paid on this debt is increasing with every major issuance.

The main risk is focussed on the order books of the four main cloud computing providers (Microsoft, Amazon, Alphabet and Oracle), who account for around USD 2.3 trillion in as-yet-unrealised contractual revenues. A significant part of this sum is owed by two private companies that are not very transparent: OpenAI and Anthropic. Anthropic is in the better position, with a positive adjusted operating profit for its annualised turnover of USD 65 billion, while OpenAI is set to burn through more than USD 200 billion in cash by 2029. Elsewhere, Moody’s is stressing that the system is relatively circular, which hides real demand: chip manufacturers and cloud providers are investing in the labs that are buying their products, with Alphabet going so far as to guarantee the managers’ leases on Anthropic’s data centres.

The bond market will continue to soak up the debt that is being issued, but at a revised price for each major issue. The calendar is key here: the contracts signed in 2025/26 will start to be billed in 2027/28, just when capex growth is set to slow sharply. Whether 2027 will be a tough year or a simply busy one will depend on the ability of OpenAI and Anthropic to pay their bills; if they fail to do so, renegotiations should be expected rather than defaults.

Microsoft, Amazon, Alphabet and especially Meta, whose AI is already profitable thanks to advertising revenues, seem to be in the best position: they are being asked to absorb a bond-issuance programme, not a credit problem. Oracle, which is more indebted and half of whose order books depend on OpenAI, along with the neoclouds, which rely on a small number of clients, have, for their parts, to absorb both a significant amount of the debt being issued alongside a real credit risk.