AI's biggest spenders feel the credit squeeze

The primary corporate bond market is struggling to digest a flood of debt from the companies spending most on artificial intelligence, and institutional credit portfolios are full. Credit default swap (CDS) spreads for the big technology names, Oracle above all, have widened sharply over the past month.

The turn has been quick. In September Oracle sold $18 billion of bonds and its shares soared after it disclosed a surge in contracted future revenue, much of it tied to a multi-year computing deal with OpenAI. Since then Meta has priced its $30 billion deal, other hyperscalers have followed, and the question in credit markets has shifted from who wants AI exposure to how much of it one portfolio can hold.

Too much tech paper

Oracle, Meta, Microsoft and Amazon together issued more than $100 billion of debt in a single quarter to pay for data centre construction and GPUs. Institutional credit managers have now hit the statutory limits on how much they may hold of any single technology issuer. To make room for new deals, bond desks are selling existing corporate paper in the secondary market, which pushes spreads wider across investment-grade tech.

The supply is unlikely to slow. At their results at the end of October, the largest cloud operators all said they would spend more on capital projects next year than this, and none of them proposed to pay for it entirely from cash. Meta's order book of close to $125 billion showed that demand for a single deal can still be enormous. The difficulty is that every fresh deal needs buyers who have room left, and each one leaves less room for the next.

Why Oracle leads the sell-off

Oracle is the weakest of the big spenders on paper. Unlike Meta or Microsoft, it came into the AI boom with substantial debt from earlier acquisitions and buybacks, its free cash flow has turned negative as capital spending climbed, and much of its future revenue depends on a single customer that is itself losing money. Its credit rating sits in the lower half of investment grade, where any further slide towards junk would matter to holders bound by rating rules.

The CDS market is where that anxiety shows first. A default swap is insurance against non-payment, and it can be bought without owning the bonds. Banks that have lent to data-centre projects, and bondholders who cannot easily sell, can hedge there instead, so rising demand for protection lifts the price even if no one expects a default soon. Widening CDS spreads then feed back into new-issue pricing, because buyers use them as a gauge of what compensation to demand.

A proxy for the whole trade

Some of Oracle's move reflects other companies' risk. Default swaps on the cash-rich giants trade thinly, and protection on a company with Microsoft's balance sheet offers little payoff, so investors who want to hedge exposure to the AI build-out as a whole have turned to Oracle, the most indebted large name tied to it. That makes Oracle's CDS a gauge of sentiment towards the entire trade. It also means the spread can overshoot what Oracle's own finances justify, and can snap back just as fast when the mood improves.

Borrowing costs more

The overhang of supply is raising the marginal cost of capital for the AI build-out, perhaps for good. Issuers that could once sell ten-year debt at fifty basis points over Treasuries are now paying spreads of more than 120 basis points.

The extra cost falls unevenly. The cash-rich giants can slow issuance and fund more from operations, so a wider spread is an irritation for them. Smaller operators that rent out GPU capacity, such as CoreWeave, borrow in the high-yield market and depend on continuous access to it; for them wider spreads can decide whether a project is built. New buyers are the winners, locking in better yields on paper from strong issuers, while existing holders absorb the mark-to-market losses.

Other borrowers pay too. Investment-grade issuers outside technology, from utilities to industrial companies, compete for the same pool of insurance and pension money, and a crowded calendar forces all of them to offer bigger concessions. Utilities are a particular case, since many are borrowing heavily themselves to build the power supply that data centres need.

Wide only against a tight start

The counterargument is that the move is small in context. Overall investment-grade spreads touched their tightest levels since the late 1990s this autumn, and the Federal Reserve has cut rates twice since September, so borrowing remains cheap by historical standards. A widening from very tight levels could be a healthy repricing rather than the start of a squeeze. That is a fair reading for the strongest names, but it does not fit Oracle, whose spreads have moved far more than the market's.

The late 1990s offer the obvious parallel. Telecom operators became the largest issuers in the corporate bond market while building fibre networks, and when investors grew uneasy about the returns, spreads on the heaviest borrowers widened well before any default. The signals to watch now are Oracle's quarterly results in December, any change in its rating outlook, and the size of the concessions demanded on the next large tech deal. Bond investors will keep financing AI infrastructure, but they now want to be paid properly for the risk. Wider spreads will make every new data centre dearer to fund, and that should bring some financial discipline to Silicon Valley's race for compute.