The AI Trade Moved From Stocks to Bonds — and That's Why the Long End Broke

If you want to understand why 30-year government bond yields hit their highest levels in nearly two decades this month, you have to stop looking at central banks for a minute and look at data centres.

The AI boom has quietly changed what it is. For its first two years it was an equity story funded by the operating cash flow of the most profitable companies on earth. In 2026 it has become a credit story.

The scale of the shift

The numbers are stark. The five biggest US hyperscalers used to issue roughly $28 billion a year in corporate bonds between 2020 and 2024. In 2025 that jumped to about $121 billion. In the first five months of 2026 alone they issued around $159 billion — already past the entire prior year.

Beyond the giants, the borrowing is broader and stranger. Wednesday's Nebius deal is a good example: $4.5 billion of convertible notes, split across 2030 and 2034 maturities, explicitly to fund data centres and GPUs.

Convertibles are worth understanding because they are everywhere in this cycle. A convertible bond lets the issuer borrow cheaply — sometimes at coupons far below straight debt — in exchange for the option to convert into equity if the share price rises past a threshold. The company gets cheap money now. Existing shareholders get diluted later if things go well.

Layer in private credit — which BIS research put at over $200 billion outstanding to AI-related companies, from essentially zero a few years ago — and lease structures that sit off balance sheets, and the visible bond market is only part of the picture.

Why this shows up in your bond fund

Here is the connection nobody makes on the evening news.

Investors who buy long-dated bonds — pension funds, insurers, sovereign funds — have a roughly fixed appetite for duration risk. Governments are already flooding them with supply: the US Treasury had to double its long-bond buybacks this week after a buyers' strike since late June, and European issuers priced a record post-summer week.

Now add hundreds of billions of dollars of AI-related corporate paper competing for the same limited pool of duration demand.

When supply rises and demand does not, the price falls and the yield rises. That is most of what has happened at the long end of the curve this year, and it is why a JPMorgan credit strategist has warned that bond portfolios which historically traded with rates and banks are now going to trade with technology companies' performance.

Read that again if you own a bond fund for safety. Your "boring" allocation may now carry more correlation to the AI trade than you intended.

The warning sign that is already flashing

The clearest early indicator is not price, it is demand coverage.

In February, orders for hyperscaler bond deals covered nearly five times the amount on offer. By July, coverage had slipped below two times. Deal mix explains part of that, so treat it as directional rather than definitive. But the direction is one way: buyers are getting choosier, which means issuers will eventually have to pay more.

Paying more for capital changes the arithmetic of a data centre. These projects are underwritten on decades-long asset lives and assumed future AI revenue. If the cost of funding rises while revenue arrives more slowly than modelled, the gap has to close somewhere — capex cuts, equity issuance, or credit stress.

What this means for ordinary investors

1. Your S&P 500 fund and your bond fund may be exposed to the same thesis. Equity concentration in AI names is well documented. Credit concentration in AI-linked issuers is newer and less visible. Diversification across asset classes assumes those asset classes fail for different reasons. That assumption is weaker than it was.

2. Off-balance-sheet structures make it hard to see who bears the risk. Analysis of hyperscaler lease commitments suggests large obligations that do not appear as debt on reported balance sheets. When financing migrates to private credit and special purpose vehicles, public disclosures no longer tell the full story.

3. This is not a prediction of collapse. The hyperscalers are among the most cash-generative companies in history and are not in financial difficulty. The point is narrower and more useful: the marginal dollar funding AI is now borrowed, and borrowed money behaves differently from retained earnings when conditions tighten.

4. It gives you a concrete thing to watch. Follow bond order coverage ratios on big AI deals, and follow whether convertible issuance keeps accelerating. Those will tell you the market is straining well before any equity index does.

We wrote earlier this year about the circular financing inside the AI boom — companies investing in each other and then buying from each other. Debt financing is the next chapter of that story, and it is a more consequential one, because credit markets transmit stress faster and further than equity markets do.

Frequently asked questions