Bond investors warn AI buildout could hit concentration limits

In short: As big tech companies borrow more to pay for AI data centres and chips, bond investors say they may hit limits on how much of one company’s debt they can hold.

What's going on

Building the computers, data centres, and power supplies needed for AI is expected to cost huge sums. Goldman Sachs recently raised its estimate for AI infrastructure spending to $8.5 trillion from 2026 to 2031, and expects about 35 per cent of that, close to $3 trillion, to be funded with debt.

S&P Global Ratings also says spending forecasts keep rising. It estimates that the biggest six “hyperscalers” (very large cloud and internet companies) could spend more than $7 trillion from 2025 to 2030. S&P expects their reported debt could rise to about $1.2 trillion by 2029, and their adjusted debt could approach $2 trillion. Adjusted debt is a broader measure that also counts big long-term commitments like leases (similar to counting not just your mortgage, but also a long car lease and other fixed payments).

The worry is not that these companies cannot pay. It is that many bond managers are not allowed to put too many eggs in one basket. Large investors often limit exposure to any single borrower to around 1 to 2 per cent of a portfolio. If a handful of AI companies issue a lot more bonds, some funds may be forced to stop buying, even if the interest rate offered goes up.

What to watch

Investors told the Financial Times that credit spreads, which are the extra interest companies must pay compared with safer government bonds, could rise simply because portfolios are getting “full.” If that continues, AI companies may need to rely more on other funding options, such as selling shares, using banks, or creating more specialized funds.

Source: Financial Times