BIS says the AI data-centre buildout is becoming a credit story

The Bank for International Settlements wants the AI infrastructure story read less as a triumph-of-compute narrative and more as a financing problem. In Bulletin 120, the central-bank forum says the next phase of the data-centre buildout will rely less on operating cash flow and more on debt, private credit and off-balance-sheet structures that can look tidy until refinancing suddenly gets harder.

Not a crash call. A credit call. The bulletin does not predict an imminent repeat of 2008, and it does not argue that AI demand is fake. Its point is narrower, and sharper. Once capital expenditure pushes beyond what hyperscalers and suppliers can fund internally, the story stops being only about chips, cloud demand and market share. It becomes a question of who is supplying the capital, on what terms, and how much risk is being parked in vehicles, guarantees and lease commitments that sit outside the cleanest headline numbers.

That is where the market mood starts to split. The regulator-policy view is that the system is carrying more debt. The analyst view is that investors still seem happy to price the buildout as secular growth. In a Reuters Open Interest analysis, Mike Dolan made essentially that point: if returns disappoint, financing may tighten before belief in the AI growth story does.

The buildout is getting too big for cash flow

Scale is the first warning. Reuters has put 2026 AI capital spending by the five biggest hyperscalers at about $1 trillion, with cumulative outlays reaching $7.6 trillion by 2031. At that size, the issue is no longer whether demand for compute exists. It is whether the funding mix can keep expanding without changing the risk profile of the whole stack.

The BIS answer is blunt. Its economists say the buildout will have to lean harder on borrowing because internal cash generation is no longer enough for the size of the investment pipeline.

“The size of anticipated investment needs will require firms to shift the source of financing from operating cash flows to debt, with private credit playing a rapidly increasing role.”
Source: BIS bulletin authors, Bank for International Settlements

That answers the main analyst question hanging over the sector: when does capex stop signalling strength and start signalling dependence on outside financing? Not on a single date. At the point when cloud landlords, utilities, chip buyers and model developers all need markets to stay open at once.

Recent deals show that shift in real time. Alphabet has sought fresh equity to fund its buildout. Amazon tapped banks for another $17.5 billion. Nvidia moved toward its first large debt sale since the current AI cycle took off. None of that is a distress signal by itself. Collectively, though, it shows even the strongest names are broadening the funding base.

The fragility sits in the plumbing, not the headlines

Here the credit-sceptic perspective matters more than the hype cycle. Fast Company reported that Oracle had about $160 billion in outstanding liabilities, with $133 billion tied to the AI buildout, while hyperscalers were carrying roughly $662 billion in off-balance-sheet commitments. The same report pegged private credit outstanding at about $3 trillion by late 2025.

That is the plumbing. Leverage in this cycle is not always showing up in plain-vanilla corporate bonds. It can sit inside project vehicles, sale-and-leaseback structures, credit enhancement, long-dated supply agreements and sponsor guarantees. Those arrangements are not inherently unsound. The problem is that they can make the system look less correlated than it really is.

“Disappointment in returns could trigger a sudden pullback in financing and turn the capex boom into a protracted investment bust.”
Source: Mike Dolan, Reuters Open Interest

The Semafor report on Anthropic sharpened the point. One of the world’s most important model developers, it argued, still needed a blue-chip backstop to enter credit markets comfortably. That partly answers the sceptic’s question about who wears the risk first. Not just young AI companies. Also the supposedly safer structures that depend on anchor tenants, sponsor support and lenders believing those names will always refinance.

Debt markets may blink before equity markets do

Equity investors are still treating the AI buildout as a secular growth story. Credit markets are beginning to treat it as a financing regime. That mismatch is what makes the BIS intervention notable. The institution is not trying to call the top in AI demand. It is pointing to a valuation gap between the optimism embedded in share prices and the colder assumptions lenders eventually apply to duration, collateral and refinancing risk.

“the financing of the AI boom also looks increasingly reliant on debt and complex funding structures across the supply chain.”
Source: CNBC coverage citing Reuters, CNBC

The live examples are not abstract. Schneider Electric has leaned on buoyant AI demand to sell debt. CoreWeave took its funding story into the European junk-bond market. Each deal can work on its own terms. Taken together, they show how quickly the AI trade is spreading into every layer of capital markets, from equity raises to higher-risk credit.

That helps answer the analyst’s second question about what reprices first. Probably not a dramatic collapse in tech equities, at least not initially. More likely a smaller and nastier change in lender appetite: wider spreads, tougher covenants, shorter duration, or more insistence on sponsor support. That would still matter because the buildout is being planned on the assumption that capital remains abundant.

The fallout would not stop at hyperscalers

The infrastructure-insider view is the most practical one. Risk now extends well beyond model labs and cloud groups. Data-centre landlords, power suppliers and equipment vendors are all tied to the same build cycle. When the cycle is hot, the financing looks diversified. When it slows, the correlations can become obvious very quickly.

The Financial Times recently tied booming AI electricity demand to utility consolidation. CNBC’s report on Digital Realty and Blackstone’s Virginia data centres showed private capital moving deeper into landlord financing. And Sherwood’s Oracle earnings coverage pointed to expectations that hyperscalers will need tens of billions more in debt or equity to sustain the capex push.

So the BIS warning does not stay inside the tech column. If one part of the chain slows, the effect will not be confined to server orders. It can hit lease assumptions, utility investment plans, power pricing and the appetite of private lenders that financed the boom on the assumption of near-continuous demand growth.

For enterprise buyers, including Australian organisations that increasingly depend on rented cloud and AI capacity rather than owned infrastructure, that is the practical takeaway. The race to supply more compute is not just a product roadmap. It is a capital-markets project. If financing conditions tighten, deployment schedules, pricing and supplier behaviour can all shift even if demand for AI tools remains intact.

What the BIS is actually saying

The cleanest reading of the bulletin is also the least sensational one. The BIS is not arguing that AI is a mirage, or that a financial accident is certain. It is arguing that the funding model behind the AI data-centre boom is taking on more debt, becoming more opaque and getting more intertwined with non-bank credit.

Hard questions follow from that. Who funds the next data hall? Who guarantees the lease? Who wears the refinancing risk if utilisation undershoots? What happens to the rest of the chain if lenders decide the story now looks more like infrastructure finance than software growth?

Those are central-bank questions, not hype-cycle questions. Right now, they also look like the right ones.