For two years the European argument about data centers was about power. Whether the grid could carry the load, whether a substation could be built before the servers arrived, whether the continent had enough electricity to keep pace with the American build. The question now circulating in Frankfurt is narrower and less comfortable: who holds the paper.
The European Central Bank is assessing the exposure its banks carry to data center lending and is weighing a dedicated review of it, according to market reports on Nov 7. No formal decision has been announced and the scope of any exercise has not been made public.
The exposure is not small. European banks have already arranged tens of billions of dollars in debt financing for AI data centers, and two problems surfaced together in the autumn: buyers were hard to find, and tenors stretched.
Both are symptoms of the same thing. A data center loan is normally syndicated. The originating bank underwrites the whole facility, then sells slices to other institutions, keeping a fee and a small retained position. When the slices do not clear, the loan stays where it was written. When tenors stretch, it means underwriters could not place the paper at the maturity the borrower wanted and had to hold longer duration on their own books.
What sits behind the loan is a shell company whose assets are a lease and a power contract. The building is worth what the tenant pays. That structure is why the loans looked safe enough to write in volume. A hyperscaler with an investment-grade rating signing a fifteen-year lease produces cash flow that behaves like infrastructure, and infrastructure lending is a business European banks understand.
The comparison only goes so far. A toll road has a forty-year useful life and a regulated tariff. A data hall built around a particular generation of accelerator has an economic life that depends on what the tenant does next, and the tenant is a company whose capital spending plans change every quarter. The steel does not depreciate on the same schedule as the chip inside it.
Supervisors tend to look at three things in a portfolio like this. Concentration, because a new asset class in Europe is usually held by a handful of banks and lent to a handful of borrowers. Tenor mismatch, because the loan runs longer than the lease that supports it. And collateral valuation, because nobody has watched a data center sell in a downturn.
The mechanics of a review are unglamorous. The central bank asks a sample of banks to break out their exposure by borrower, by country and by stage of construction, then compares the answers. Where banks classify the same kind of loan differently, the exercise has produced a finding before it has produced a number.
None of that requires a finding of imprudence. A review can end in guidance, in better documentation standards, in a request that banks stress the assets against a fall in rental rates. Data center lending is small next to European mortgage books, and a supervisor with a broad mandate has larger things to worry about.
The reason to look anyway is the absence of a loss history. Every credit model in the asset class is built on assumptions rather than observed defaults, and the assumptions were written during a period when demand for compute only went up. A review is a way of asking what the model does when it does not.
The transmission to the real economy is direct and fast. Construction is financed before it is built, and a project underwritten at a fixed rent roll has little room for a hundred basis points of additional cost. When the cost of debt moves, the developer does not shrink the building. It moves the start date. A delayed start date reaches the general contractor, the switchgear order, the transformer queue, and eventually the tenant’s capacity plan.
There is a second-order effect that the review itself may produce. If supervised bank lending becomes slower or more expensive, the financing migrates to private credit funds, which are less visible to a central bank and charge more for the same risk. A supervisor looking for the exposure can push it into a corner where it cannot be counted, which is a familiar outcome in European credit markets and rarely the intended one.
The banks also have a defense worth taking seriously. Their data center books are tiny as a share of total lending, the borrowers are large and rated, and the loans are secured by physical assets in jurisdictions where the lender can enforce. If the review is calibrated to the size of the exposure, it will find very little.
If it is calibrated to the concentration and the missing loss data, it will find something more interesting. The question Frankfurt is asking is which of those two exercises it wants to run, and the answer will shape how quickly the next tranche of European capacity gets built.
