AI infrastructure borrowing forecast at $420bn in 2027 as bond buyers demand more
Goldman Sachs data cited by Reuters projects record gross hyperscaler issuance next year. Investors are asking whether data-centre returns justify the scale and concentration of borrowing.
The short version
- What happened
- The $420 billion figure is Goldman's forecast for gross debt issuance in 2027, around 60% above its 2026 estimate.
- Who it affects
- Pension funds and savers with corporate-bond exposure may see concentration in a small group of AI infrastructure issuers.
- What remains unproven
- Goldman Sachs' estimate is forward-looking and can change with interest rates, spending plans and market access.
Keep in mind
Goldman Sachs' estimate is forward-looking and can change with interest rates, spending plans and market access.
Reuters' interviews capture selected investors' views rather than a complete survey of bond buyers.
Source analysis · 1 linked source. Primary analysed source date: 22 September 2026.
Read the full report · 3 min The evidence, methods, implications and limitations
What the forecast measures
Reuters reports Goldman Sachs data projecting that gross bond issuance by hyperscalers could reach $420 billion in 2027, up roughly 60% from its 2026 estimate. Gross issuance counts new borrowing over a period; it is not the outstanding debt stock, a measure of spending already completed or a forecast of losses. The distinction matters when a large number is repeated as if it were an incurred bill for AI.
The borrowing would help finance chips, data centres and associated infrastructure. Credit investors interviewed by Reuters did not describe an immediate fear of default among the largest issuers. They focused instead on the volume and unpredictability of supply, portfolio concentration and incomplete visibility into returns on invested capital.[1]
Why bond pricing is changing
When many issuers seek funding for similar projects at once, bond buyers can ask for higher yields or new-issue concessions. Reuters points to a split between AI-linked debt and some traditional corporate issues that attracted stronger demand. The comparison describes market conditions at the time; it does not prove all AI investments are uneconomic or that every issuer faces the same terms.
A data centre's revenue depends on tenant demand, utilisation, electricity access and the useful life of equipment. Chips can depreciate quickly while buildings and power agreements last longer. A strong balance sheet can absorb a disappointing project, but lenders and investors still have reason to ask how each investment becomes durable cash flow.[1]
Evidence to follow
The most useful disclosures separate committed capital expenditure, financed versus internally funded spending, contracted capacity, utilisation and returns. Investors also need to distinguish firm customer commitments from optimistic forecasts of future model use. Bond terms, maturities and interest costs show how much flexibility an issuer keeps if demand changes.
Local communities face another dimension: grid expansion, water use and land decisions may be made before commercial returns are clear. Public infrastructure costs should be visible separately from private financing. The market signal today is greater selectivity, not a conclusive verdict on the long-run economics of AI.[1]