Positive feedback loops with capacity expansion announcements can turn an AI arms race into overcapacity.

This is the third article in a four-part weekly series. This series is based on my new 12,000-word research paper, The Economics of AI Infrastructure, and the financial model used in that analysis. The first article AI may succeed spectacularly while infrastructure investments fail distinguished demand growth from capital recovery. The second article last week AI is not one business. It is two. argued that AI infrastructure and frontier-model platforms are different businesses. This article focuses on utilization and investment feedback loops.

AI infrastructure commentary often focuses on large gigawatt construction projects, including supply of chips, electrical power and cooling. How fully the resulting capacity will be used is rarely questioned in face of surging demand right now.

The utilization rate could become the most important number in the economics of AI data centers.

I use a central illustrative assumption of 75% for AI data center utilization in my financial analysis. This is not an empirical estimate. It is a modeling assumption tested and varied in my sensitivity analysis. This shows how vulnerable per unit costs are when a large share of the total cost is fixed or quasi-fixed.

Electricity price matters, but it is not the largest cost component. In a gigawatt-scale AI data center, annualized recovery of silicon, servers and high-speed networking can dominate total annual cost because those assets are expensive and depreciate quickly. A partially utilized facility still incurs most of that capital recovery burden.

That is why utilization changes everything. If expected utilization is 75% but actual utilization falls much lower, the fully absorbed cost per query, token or useful AI task rises sharply. The facility may still be technically useful. It may even remain a money-spinner for a second owner buying assets for pennies on the dollar. But first-owner financial return assumptions can fail.

This dynamic is familiar from fiber networks, airlines, semiconductor fabs and other high-fixed-cost industries. A little underutilization can be manageable. Persistent underutilization can be devastating. A key reason why the cellular tower colocation sector has performed well over decades is that towers are typically constructed only when an anchor tenant is secured with a long-term contract. This ensures a minimum return on investment.

There is another risk now visible in AI: positive feedback loops in capital spending.

In a rational commodity supply market, large capital investment by one competitor can serve as a warning to others. If a petrochemical company announces a massive new plant, rivals may reconsider building another when there is a common understanding of likely total market demand. That is a negative feedback loop. Supply increases restrain additional supply increases.

AI currently seems to exhibit the opposite phenomenon. One company’s enormous AI spending appears to be interpreted by others as evidence that they too must spend more to increase their capacity and capture their share of demand growth. AI is considered winner-take-most, and being late or small is existentially dangerous. Investment begets investment.

That arms race may be rational for individual companies, especially if they are defending search, cloud, software, advertising or social-media franchises. But it can still be damaging at the supporting infrastructure level. Supply and demand cannot both be infinite. At some point, somebody must assess how much AI data center capacity is actually being committed and financed, what are the contingencies, how likely it is to be completed, where it is located, what pricing and revenues it requires.

That is why I am coupling my financial model with the tracking of data center capacity expansions, associated financing and demand indicators. Public announcements already provide fragments: gigawatts, capex, power agreements, financing packages including those supported by equipment vendors, hyperscaler commitments and model-provider contracts. Bringing those into a consistent framework would be one step toward being able to meaningfully forecast AI infrastructure demand growth and pricing over investment lifecycles of at least five years.

This matters because today’s market signals are noisy. Long-term contracts may be tied to equity stakes, supplier relationships or strategic commitments. Vertically integrated Big Tech companies may absorb infrastructure operation losses with broader economic considerations in their platform businesses. Neoclouds and independent infrastructure owners may not have the same flexibility.

The lesson from telecom infrastructure is not simply that overbuild happens. It is that capital markets can fund too much capacity before pricing and utilization settle into a sustainable pattern.

AI may avoid that outcome if power, water, permitting and grid constraints prevent over-supply. Scarcity can protect returns. Long-term contracts can also help. Sovereignty requirements may strengthen some regional positions.

But investors should not assume that rapid demand growth alone solves the problem. How rapid is rapid enough? What pricing is required? Utilization and capital recovery need to be modeled, and tracked. Plans and assumptions must be challenged.

The fourth and final article in this series next week asks where the value ultimately accrues: infrastructure owners, frontier-model platforms, application developers, chipmakers or end users.

Keith Mallinson, founder of WiseHarbor, has more than 30 years of experience in the ICT sector as a technical, commercial and financial research analyst, consultant, and testifying expert witness.