What to Know
- AI infrastructure investment may be diversified by asset but concentrated around shared adoption assumptions.
- Interconnected AI financing, data centers, chips and power infrastructure can amplify financial risk.
- Slower AI adoption could propagate risk across lenders, utilities, cloud providers and investors.
Consider a well-leased office tower. On its upper floors sits a law firm; below it, an accounting practice; and on the lower floors, a management consultancy. To the owner and the bank that financed it, this is a comfortable arrangement, because the tenants do not rise and fall together. There is enough diversification to survive differing economic cycles.
Now consider the infrastructure being built for AI – data centers, chips, power lines, financing and the contracts binding them together. At the level of any single transaction, it looks diversified. A chip sale is not a cloud lease; a cloud lease is not a construction loan.
Appearances can be misleading. The tower is diversified because its tenants depend on different things. The AI buildout may appear diversified, because its participants increasingly depend on the same thing: a shared expectation about the future economics of AI adoption. What appears diversified at the transaction level may be concentrated at the assumption level. This is assumption concentration: a portfolio that is diversified by asset, tenant and counterparty, but not by the premise on which each position depends.
This is not the bubble argument. The concern is not that the shared expectation is wrong, but how the system would respond if that expectation changed.
Following One AI Ddollar
To see how a shared assumption embeds itself, follow a single dollar of expected AI demand as it moves through the system.
It begins as a forecast inside a model developer. On the strength of that projection, the developer commits to a cloud provider for an enormous quantity of capacity, often years before the revenue to pay for it exists. The cloud provider contracts in turn with a data center developer to build the physical capacity. The data center developer borrows, and the lender takes comfort from the chain above it: a signed cloud contract, behind which stands a credible model developer. The loan funds construction, which requires power, and so a utility commits capital to generation and transmission with operating lives measured in decades. The chips that fill the building come from a semiconductor manufacturer, whose own production and investment plans are scaled to the same expected demand. And investors supply equity, underwriting valuations that presume the forecast will be borne out.
Participants may believe they are underwriting different risks while depending on the same economic outcome. The dollar of demand has not been diversified by its passage through the system. It has been counted again at every link. Call it the daisy chain.
A careful reader may object that some of these assumptions are binding obligations. Contractual demand reduces propagation risk; it does not eliminate it. A contract is only as reliable as the counterparty behind it and the end-use economics that make honoring it rational, and those economics rest on the same adoption premise with which the chain began. The daisy chain does not disappear when each link is properly documented. It runs one layer deeper.
Those contractual protections are the ordinary, sound machinery of project finance, refined over decades to allocate risk to the party best able to bear it. What no single instrument is built to reveal is that the risks it has so carefully separated may, this time, be the same risk wearing different paper. The structures are working exactly as designed. The question is whether what they were designed to assume still holds.
