- The AI Race Is Entering a New Phase: Who Pays, And Who Can Repay?
The artificial intelligence infrastructure boom is moving into a more debt-intensive phase. Goldman Sachs Research estimates that global hyperscaler debt issuance could rise from $250 billion in 2026 to $420 billion in 2027, an increase of 68.00%.
The shift puts a sharper question before investors and company boards: can the revenue generated by AI infrastructure justify the borrowing used to build it?
The forecast concerns direct debt issuance by hyperscalers, the large technology companies that build and operate extensive cloud and computing infrastructure. It is not a measure of all borrowing linked to AI.
Goldman estimates aggregate capital expenditure across compute, data centres and power could reach $7.6 trillion between 2026 and 2031. It also says hyperscalers represent about 40% of credit raised across the broader AI ecosystem.
Those figures show both the scale of the investment opportunity and the difficulty of assessing its eventual returns. Data centres, computing equipment and power infrastructure require large sums before they produce revenue.
The spending may support future growth, but the debt used to finance it creates fixed repayment obligations regardless of whether customer demand or AI monetisation develops as quickly as expected.
For now, the borrowing forecast does not by itself point to imminent financial distress. Goldman has noted that hyperscalers remain highly rated and relatively lightly leveraged, giving them room to add debt.
The issue is how much new supply bond markets can absorb, and at what price, without investors becoming overexposed to a single technology theme or a small group of issuers.
That distinction matters. A company can have the capacity to borrow and still face pressure if it takes on debt faster than its investment produces cash. The risk is especially relevant when the infrastructure being financed is long-lived, while customer demand, technology standards and commercial models can change quickly.
Goldman’s analysis highlights the timing gap. Hyperscalers’ capital spending is approaching their operating cash flow, while the period between investment and monetisation can range from three months to two years.
Goldman expects debt to finance about 35% of hyperscaler capital expenditure in 2027, up from roughly one-third in 2026. The estimate reflects an expected financing mix, not a guarantee that the spending will produce returns at the same pace.
This is where capital discipline becomes more than a question of limiting borrowing. It means testing whether each project has a credible route to utilisation and revenue, whether the timing of expected cash flows matches the debt repayment schedule, and whether the company can still meet obligations if construction is delayed or demand disappoints.
Goldman has also estimated that project finance and data-centre transactions could add another $300 billion in 2027, beyond direct hyperscaler issuance.
That separate estimate underlines why the $420 billion figure should not be mistaken for the full credit exposure to AI infrastructure. Debt is being raised across a network of technology companies, data-centre projects and other financing structures.
Investor appetite is unlikely to be determined by the size of the debt alone. It will also depend on the structure of each deal, the borrower’s existing balance sheet, the maturity of the debt and the clarity of the project’s cash flows. Goldman’s credit-market discussion described the coming period as requiring a “more nuanced conversation about where to actually take this risk and at what price.”
That selectivity could affect the cost of capital. If investors demand higher returns for taking on concentrated exposure, companies may have to pay more to issue debt.
Higher financing costs can reduce the expected return on a project, particularly if power connections, equipment delivery or construction take longer than planned. For a project financed on optimistic assumptions about utilisation, even a modest delay can complicate repayment.
The risks are not confined to the largest technology groups. A wider field of data-centre developers, infrastructure operators and other AI-linked businesses is seeking capital.
Some may have less diversified revenue and fewer financing options than the hyperscalers. If public markets become less receptive, these borrowers may turn to private credit, project finance or other structures.
That can broaden the sources of funding but may also make it harder to compare total exposure across the sector.
For company boards, the discipline test should start before a loan is issued. Borrowing should be linked to realistic operating cash flows, with stress tests for lower utilisation, slower customer adoption, higher energy costs and construction delays.
Firms should also consider how much investment can be funded internally, whether equity or partnerships would better distribute risk, and how much debt the business can service without sacrificing strategic flexibility.
For investors, the relevant question is not simply whether AI is likely to grow. It is whether a particular borrower can convert its infrastructure spending into cash flows sufficient to service the debt.
A project can be strategically important and still deliver returns later than lenders expect. A strong balance sheet can cushion that delay, but it does not remove the need for repayment capacity.
The $420 billion forecast therefore marks a shift in how the AI boom will be financed, rather than a verdict on its eventual success. Borrowing can accelerate infrastructure development and support innovation.
But as debt takes a larger role, the quality of underwriting, the transparency of project economics and the fit between repayment schedules and cash generation will become central to the durability of the expansion.
