Why a Bigger Boom Can Tip into a Deeper Bust

The AI infrastructure build-out becomes fragile at the point where capacity committed before uncertainty is resolved raises the productivity needed to make the plan pay. This chapter follows that transition in the BIS model. The practical distinction is between the chance of crossing into a bust, the damage conditional on crossing it, and the expected economic value of the build-out; those are related outcomes, but they are not interchangeable.

A larger build raises the break-even bar

The model has a simple time structure. In the first period, each coalition commits an initial amount of capacity. In the second period, a productivity draw arrives. It can represent faster adoption and stronger capabilities when it is high, or disappointment when it is low. The coalition then decides whether to deploy more capacity. In the third period, the contest is resolved and debt is repaid from the realised payoff.

This ordering matters. If the productivity draw were known before any investment, a coalition could wait and avoid sinking as much capital into a weak state. But the head-start advantage makes waiting costly. The coalition commits early to secure a position, then discovers whether the technology and its demand can support the commitment. The early investment is not automatically a mistake; it is the price of competing for a lead. The fragility comes from the interaction between that price and diminishing returns.

Let the initial commitment be I₁. After the draw, the coalition considers period-two investment. In a good state, it can deploy from its retained circular budget and borrow against marked-to-market collateral. It continues until the marginal return on the next unit falls to the cost of capital or a financing limit binds. In a bad state, the marginal return is already below the cost of capital. Further deployment stops.

The boundary between these states is an endogenous continuation threshold. In the paper’s notation, the draw must exceed a threshold ρ̄ for period-two deployment to be worthwhile. The threshold rises with I₁. The reason is not that a larger data centre mechanically makes the model less capable. It is that concave sector demand makes the additional output required to cover the cost of a larger early commitment increasingly difficult to obtain. A bigger initial build leaves a smaller margin for disappointment.

This is the central endogenous-fragility result. The boom changes the system’s exposure before any bad news arrives. If the coalition had built less, a given productivity draw might have supported continued expansion. After a larger commitment, the same draw can fall below the break-even threshold. The shock has not become worse; the investment plan has made it insufficient.