The bond-market advantage behind AI’s biggest builders

The AI boom has divided the technology industry in an unexpected place: the price companies pay to borrow. In research published in September, Bank of Italy economists Marco Albori and Andrea Zaghini find that the largest cloud groups secured better bond financing relative to other issuers after ChatGPT’s arrival. Their biggest estimated advantage was against software firms.[1]

That is a consequential distinction for an industry buying buildings, electricity systems and computing equipment years ahead of the revenue those investments are meant to earn. The cost of financing this interval influences which businesses can expand and how much cash they must devote to servicing the expansion.

The comparison is between borrowers

The study examines US-market bond placements from January 2020 to February 2026. It compares Alphabet, Amazon, Meta, Microsoft and Oracle with other issuers before and after 30 November 2022, adjusting for bond and issuer characteristics and time-varying country and currency effects. Reweighting makes the comparison bonds more similar on observable characteristics.[2]

The main table estimates a post-ChatGPT yield discount of about 29 basis points against other non-financial companies, rising to about 42 basis points against software firms. These are estimated relative gaps in yields at issuance. They are neither every company’s borrowing rate nor evidence that interest rates fell across the economy.[2]

The software comparison is statistically stronger: it clears the paper’s 1% significance threshold, whereas the broad-market discount clears only 10%. The design also depends on comparable yield trends without the AI shock. The authors examine earlier trends, finding no systematic divergence but one significant pre-event coefficient. This is an observational estimate of repricing.[2]

A small rate difference on a large sum

Scale makes the difference worth examining. The authors’ October account reports about $108 billion of issuance by the five groups in 2025, against an annual average near $20 billion in 2023–24. Their dataset ends in February 2026, so the study does not establish the discount available on a new bond today.[1]

BIS analysis supplies the financing context: much of the hyperscalers’ 2025 bond issuance had maturities beyond five years, locking in funding for construction programmes. A lower rate agreed at issue can ease that programme’s payment burden. It does not reduce the number of substations, servers or paying customers a project needs.[3]

Confidence becomes a financing input

The authors interpret the gap as investors distinguishing infrastructure suppliers from software businesses exposed to disruption. In that reading, lenders expect the cloud groups to capture more dependable future cash flows. The paper measures how those expectations appear in debt pricing; it does not establish the realised return on the data centres.[1]

Nor does relative advantage eliminate absolute risk. The March BIS analysis also observed rising prices for credit-default protection, particularly for lower-rated hyperscalers. Those contracts measure a different market from new bond placements, over a different comparison. A group can retain an advantage against software firms while lenders become more concerned about parts of its borrowing.[3]

For competition, the interesting feedback is straightforward. Confidence can lower the financing hurdle for another project; a lower hurdle can make expansion easier to sustain. The balance sheet becomes part of the contest for AI customers before the next service reaches them. Whether the resulting capacity earns enough remains an operating question that a favourable bond price cannot settle.

Report an error or suggest a correction