What the AI Boom Does—and Does Not—Tell Us About Financial Fragility
The current AI build-out presents a problem that cannot be resolved by choosing between “real demand” and “speculative excess.” This series has examined how both can exist: realised use can support a large expansion, while a winner-take-most race can make firms build too much, too early, and through financing links that magnify a disappointment. The final judgment is therefore conditional. The evidence supports a serious mechanism for endogenous fragility, but it does not provide a dated crisis forecast or a complete map of every institution exposed.
A real boom can still be structurally fragile
The integrated chain begins with demand. Major hyperscalers are reporting or guiding very large capital expenditures, and the BIS paper places major-hyperscaler capex above $700 billion in 2026, with industry expectations of several trillion dollars of AI-infrastructure investment over the following years. The independent OpenRouter study observes realised consumption across more than 400 models: 380.8 trillion tokens over January 2024 through April 2026, with weekly tokens rising from 11.4 billion to 15.6 trillion. These observations make a substantial build-out plausible. They are not evidence that every increment of capacity will earn an adequate return.
The second link is competition. A coalition that builds capacity may do two things at once: increase the total output of the AI sector and improve its own chance of controlling a dominant position. The first benefit is partly social; the second is partly a transfer from rivals. With concave sector demand, the private marginal return exceeds the planner’s return in the BIS contest model. That is the source of overbuilding. The model does not require the firms to misunderstand their own incentives. Each can be acting rationally to avoid losing a strategic position while the collective level of spending becomes excessive.
The third link is timing. Early capacity has a competitive value that later capacity cannot fully reproduce. It may attract users, developers, data, and complementary talent, and it gives a coalition a chance to compound a lead. The cost is that the commitment is made before the productivity draw is known. Waiting preserves an option to learn, but it may surrender the position that makes later investment valuable. The race therefore moves capital into the period of greatest uncertainty.
The fourth link is financing. Circular arrangements between a hyperscaler and a lab can solve a genuine hold-up problem by combining an equity stake with a future compute-purchase commitment. They can also concentrate exposure to the same underlying success. If a lab’s expected demand falls, the backer’s equity claim and the value of the compute relationship may weaken together. Debt adds a fixed repayment obligation to an asset whose resale value may fall when all sector participants retrench at once.
The fifth link is network structure. A lab failure can mark down a backer’s equity stake. If that markdown pushes the backer beyond its pledgeable-income limit, the backer retrenches and withdraws compute from another lab. The second lab can then become distressed even if its own productivity is adequate. Concentrated bridge backers lower the scale at which this cascade becomes possible.
The result is a particular form of fragility. A larger boom raises the productivity threshold that later deployment must clear. A more front-loaded and leveraged boom leaves more specialised capital exposed if the threshold is missed. A more concentrated financing network makes it easier for one firm’s balance-sheet loss to become another firm’s operating constraint.