Summary
- S&P warns that interconnected financing arrangements may make the full financial exposure of hyperscale technology companies difficult to assess.
- The agency estimates that six major US technology companies will spend more than $7 trillion on data centres and other AI investment between 2025 and 2030.
- No additional debt estimate or sector-wide rating downgrade has been announced.
S&P Global Ratings has warned that major technology companies could have more exposure to AI infrastructure financing than is apparent from their published disclosures, reflecting the increasingly complex relationships between hyperscalers, specialist computing providers, developers, and lenders.
Mark Habib, S&P’s global head of digital infrastructure, raised the concern during a roundtable in London on 19 October, as investors examined the growing cost of financing data centre construction and AI computing equipment.
He identified interconnected arrangements in which large technology companies provide financial support or guarantees for infrastructure developed by other businesses. Such structures can create obligations that are not readily visible to analysts examining an individual company’s reported borrowing.
Habib did not estimate the amount of additional debt or other liabilities potentially involved. S&P’s concern is therefore about visibility and the distribution of financial exposure, not a verified finding that particular companies have concealed a specified amount of borrowing.
The ratings agency also said it sees no present reason to change its assessments across the sector simply because bonds associated with AI infrastructure have fallen in market value.
Its central concern is whether revenue and profits generated by AI services will ultimately justify the substantial investment being committed to the infrastructure supporting them.
Large spending commitments involve multiple parties
S&P estimates that Amazon, Microsoft, Alphabet, Oracle, Meta, and SpaceX will collectively spend more than $7 trillion on data centres and other AI-related capital expenditure between 2025 and 2030.
The agency expects the group to experience negative free cash flow until at least 2027, according to the assessment discussed in London.
These are forecasts for a defined group of companies, not figures describing expenditure already incurred or cash flow reported by every company individually.
Infrastructure investment includes more than constructing server buildings. High-density computing deployments also require accelerators, networking hardware, electrical systems, cooling equipment, and the resources needed to operate and maintain them.
Because the equipment and facilities must be delivered before their full commercial value can be realised, large investment programmes may generate substantial near-term cash requirements.
Some technology companies finance and operate their infrastructure directly. Others enter into long-term leases, capacity purchase agreements, joint ventures, or arrangements involving separate owners of buildings and computing equipment.
Those structures can serve legitimate commercial purposes. They allow investors and operators to specialise in different assets while allocating construction costs, operating responsibilities, and financial risks through contractual agreements.
However, the companies participating in a project can remain economically dependent on one another even when their assets and liabilities are recorded separately.
A developer may borrow to construct a facility supported by an expected long-term customer commitment. If the customer changes its requirements or the project is delayed, the developer can remain responsible for financing costs while the anticipated revenue is deferred.
The customer’s exposure depends on its contractual obligations. An ordinary service purchase agreement, financial guarantee, lease, and direct debt obligation can produce different legal and accounting consequences.
They should not be treated as interchangeable evidence of borrowing.
S&P’s concern is that some of the increasingly interconnected arrangements may fall outside the transactions it currently rates or may not be described in enough detail for analysts to assess their full implications.
Physical delivery remains central to credit risk
Data centre financing is closely connected to the status of the underlying infrastructure. Revenue projections may depend on facilities obtaining sufficient electricity, securing permissions, completing construction, and installing computing equipment within contractual deadlines.
Delays can extend the period during which capital is committed but the associated operating revenue has not yet begun.
Electrical infrastructure creates particular uncertainty. A project may require new substations, transmission connections, generation arrangements, or reinforcement of existing networks before it can operate at its proposed capacity.
The timetable for those works can differ substantially from the completion date of the building itself.
Hardware investment presents another challenge because the commercial value of a computing deployment is affected by the performance and availability of newer equipment generations.
Financing arrangements therefore need to accommodate both the physical life of data centre assets and the potentially shorter commercial life of specialised computing systems.
These questions are relevant to European developments even where the underlying financing transaction involves US companies. British and European campuses frequently combine property investment, specialist infrastructure operators, equipment providers, and long-term occupier commitments.
The financial consequences of a delay will depend on which party owns the assets, which party must make payments, and whether contractual obligations continue when the facility is not yet available.
S&P has already taken individual rating action where it considers those risks material. In July, the agency downgraded Oracle to BBB-, the lowest investment-grade rating, citing AI spending and uncertainty about the path to profitability.
That decision should not be interpreted as a general downgrade of the data centre sector. Habib said recent weakness in AI-related bonds did not, by itself, establish a sufficiently significant difference between market pricing and the agency’s longer-term credit assessments.
Further rating pressure could emerge if investment continues increasing without adequate returns, or if competition weakens the revenue expected from AI services.
The latest warning leaves the amount of potentially unrecognised or insufficiently disclosed exposure unresolved. Greater transparency around guarantees, leases, project financing, and customer commitments would allow the financial risks to be assessed more precisely.
For now, S&P’s position is that the complexity of AI infrastructure financing warrants closer scrutiny, while the available evidence does not justify a blanket reassessment of hyperscaler creditworthiness.
