The AI investment case now spans two balance sheets
Artificial intelligence is often discussed as an equity-market question: which company will sell the most computing, software or semiconductor capacity? The construction programme supporting that demand also raises a credit question. Who finances the land, buildings, grid connections, cooling equipment and fast-depreciating servers before revenue covers their cost?
The answer differs across the sector. Some large technology groups can finance spending from operating cash flow. Other developers use loans, project financing, private credit or long-term leases. Treating those structures as a single AI debt bubble conceals more than it reveals. Investors need to identify the entity bearing the borrowing and utilisation risk.
Higher yields narrow the margin for error
Two facilities with identical projected demand can produce very different returns when one pays materially more for capital. Debt service starts on a contractual timetable; customer utilisation and pricing may not. Long lead times for power connections and construction make the mismatch particularly important.
Government-bond yields influence the baseline cost of financing, while credit spreads reflect additional risk. A rise in either can reduce the present value of expected future cash flows. That does not mean AI services have no value. It means the price paid for the capacity to deliver them can be too high even when end-user demand grows.
A signed contract is not the same as realised cash flow
A long-term commitment from a highly rated tenant can reduce the risk of an infrastructure project. Yet investors still need to know whether obligations are fixed, who pays for upgrades, how electricity costs are passed through and what happens if equipment becomes obsolete. Contracted demand, recognised revenue and cash available for debt service are three different measures.
Power availability creates another bottleneck. A fully financed site without a timely grid connection is not an operating data centre. Investors should distinguish announced capacity from commissioned capacity and from capacity actually utilised by paying customers.
The same assumptions can appear in several asset classes
A small group of technology companies can influence equity indices, purchase semiconductors, rent data-centre capacity and underpin the revenue forecasts of infrastructure lenders. Apparent diversification across equities, credit and real assets may therefore conceal exposure to the same underlying assumptions about AI monetisation.
The risk is correlation rather than a predetermined crash. If expectations for revenue per unit of compute fall, the impact could reach technology multiples, supplier orders, infrastructure valuations and credit pricing at the same time. Conversely, sustained utilisation and rising cash generation could validate investment that looks expensive today.
The indicators that matter beyond capital-spending headlines
The most informative measures are operating capacity, customer concentration, contracted terms, free cash flow after investment, interest coverage and the cost of replacing computing hardware. Investors should also compare project economics at different electricity prices and utilisation rates rather than relying on a single management forecast.
AI can become a large and useful technology while delivering disappointing returns to some owners of the infrastructure built around it. Separating technological adoption from financing returns is the discipline the next phase of the investment cycle demands.