Why Competition Makes Firms Build Too Much
The AI infrastructure race becomes overinvestment through a specific incentive: each hyperscaler–laboratory coalition values not only the revenue created by additional compute, but also the position it can take from rivals. This chapter isolates that mechanism. The reader should come away able to distinguish a valuable increase in sector capacity from a private decision to build more than the sector can efficiently use, and to see why the gap depends on how sharply returns diminish as capacity expands.
From productive investment to a prize contest
A useful starting point is the coalition rather than the individual company. In the BIS model, a coalition joins a hyperscaler with an AI laboratory as a joint decision-maker. That is a simplification: real firms may have separate shareholders, contracts, and bargaining positions. It is useful because it puts the question in focus. The coalitions are competing to occupy a few dominant positions in an AI market, and their decisions determine both how much capacity the sector has and who receives the rewards from that capacity.
Call the capacity that can actually contribute to the contest a coalition’s effective compute. It need not be identical to the number of processors installed. The model allows for supply frictions such as GPU rationing, power, cooling, and construction bottlenecks, and it gives earlier capacity an advantage over later capacity. For this chapter, however, the key point is simpler: each coalition chooses a scale, and the sector’s effective compute is the sum of those choices.
If total effective compute is called J, sector revenue can be represented as:
Sector revenue = productivity × technology scale × total compute<sup>β</sup>
The exponent β is below one. That notation captures a familiar economic idea: adding capacity can expand the market, but each additional unit need not produce as much as the last. Demand can become saturated. Customers may substitute among providers. Power, data, and complementary infrastructure can become bottlenecks. The sector may also discover that it has built more capacity than the most valuable applications can profitably absorb.
A concave revenue function does not imply that AI is unproductive. It says that a sector should not evaluate the hundredth unit of capacity as if it were the first. The first large deployment may unlock new services; later deployments may compete for the same customers or sit idle until prices fall. The distinction matters because the private investment decision is made by each coalition, while the diminishing returns apply to the sector as a whole.
Now add the contest. Suppose several coalitions are trying to win a small number of dominant positions. A coalition’s payoff depends partly on the total revenue that new compute creates and partly on its share of that revenue. A common way to represent the share is a contest lottery: the more effective capacity a coalition commits, especially capacity that enjoys an early lead, the greater its chance of receiving the prize. It is not necessary for the firm to believe that every extra processor creates new social value. It may invest because the processor improves its probability of capturing value that would otherwise go to a rival.
That is the business-stealing element. If coalition A adds capacity and takes users from coalition B, A’s shareholders may benefit. But the sector has not gained the full amount of A’s private return. Part of the gain is a transfer of market share. Coalition A does not normally subtract that transfer when it evaluates its own investment. A sector-wide planner would.
The distinction is easier to see with a simple comparison. Imagine that two providers are competing for the same enterprise customers. A new data-centre cluster lets one provider serve customers faster and advertise a wider model portfolio. Some of the resulting revenue comes from new AI applications. Some comes from customers switching from the other provider. The first component is a social gain, subject to the cost of the capacity. The second is a private gain from winning the contest. Both influence the provider’s decision, but only the first is an additional sector output.
The model therefore does not rely on irrational expectations. A coalition can make a defensible decision from its own perspective and still contribute to an excessive aggregate build-out. Each coalition treats rivals’ actions as given when it chooses capacity. It does not fully account for how all coalitions together push down the marginal value of capacity or how a common overbuild will affect resale prices later.
How the private return becomes a sector-wide gap
The planner’s calculation is narrower than the coalition’s. The planner asks what one more unit of compute adds to total sector revenue after accounting for diminishing returns. At a symmetric allocation, all coalitions expand together, so the planner sees the full effect of a larger total J. A coalition asks what the same unit does for its own contest payoff. The coalition receives the output it helps create, but it also receives a larger share of a prize partly taken from its rivals.
The BIS paper summarises the difference with a contest wedge:
λ = (n − 1 + β) / (nβ)
Here n is the number of competing coalitions and β is the sector’s returns-to-scale exponent. When there are at least two coalitions and β is below one, λ is greater than one. In plain language, the private marginal return is inflated relative to the social marginal return because the coalition counts a share-stealing benefit that the planner nets out.
The distinction can be stated as a marginal decision. If a coalition adds one unit of compute, its own expected contest share rises. At the same time, the denominator of the contest share—the capacity against which it is competing—changes, and the rival coalitions receive less of the prize in relative terms. The coalition counts the first effect as a benefit. It does not treat the second as a cost because the lost share belongs to other firms. A planner sees one combined sector: the new compute produces some additional revenue, but the transfer of market share does not create an additional unit of output. The wedge is therefore not a claim that firms miscalculate their private prospects. It is a claim that private prospects and sector output are different objects.
For the paper’s baseline, five coalitions and β = 0.80 give λ = (5 − 1 + 0.80)/(5 × 0.80) = 1.20. The number is a marginal multiplier inside the model, not a statement that every firm expects a 20% higher accounting return. It says that, at a symmetric allocation, the private value of an incremental unit includes a relative-position component that is absent from the planner’s calculation. The effect then passes through the investment condition, where the marginal product of capacity declines with scale. A modest wedge at the margin can produce a larger difference in the level of capital committed.
The wedge has an important boundary. As β approaches one, the sector approaches constant returns: additional compute can scale revenue without the same concavity. In that limit, the share-stealing margin disappears in the model and λ approaches one. The contest no longer creates this particular overinvestment mechanism. Conversely, when sector demand is more curved, the social return to additional capacity falls more quickly while the private desire to secure position remains. The gap becomes larger.
The number of competitors also matters, but it should not be interpreted as a simple rule that “more firms always means more waste.” The formula describes a particular winner-take-most contest. It says that, given the contest structure and diminishing returns, each coalition has an incentive to value relative position. Actual AI markets may contain several product categories, differentiated services, and complementary providers rather than one prize. Those features could weaken or reshape the wedge. The model’s conclusion is conditional: where a few firms are racing for dominant positions and sector returns are concave, private capacity decisions can overshoot the collective optimum.
The wedge is also different from a claim that the market has too many firms. In the classic entry problem, the distortion is mainly on the extensive margin: too many firms enter. Here the BIS model holds the set of competing coalitions fixed and places the distortion on the intensive margin. The same established contestants each build more capacity because a marginal dollar improves their position. That distinction matters for the AI case. Consolidating the field would not automatically make the investment level efficient; it could simply replace a contest among several builders with a different concentration problem. The model’s benchmark is a lower collectively worthwhile level of capacity, not the elimination of competition.
Nor is the planner comparing investment with zero. The planner still builds when the expected value of AI output covers the real cost of capacity and capital. The planner simply discounts the part of a coalition’s return that comes from taking a position from another coalition. This is why evidence of strong customer demand and evidence of a contest wedge can sit together without contradiction. New demand can justify a larger sector while still failing to justify every coalition’s attempt to secure the same strategic lead.
The wedge operates at the margin, but investment is a level decision. If producing effective compute is itself subject to concavity or supply friction, a higher marginal incentive has a more-than-proportional effect on the amount of spending required to reach the chosen capacity. In the model, effective compute rises with investment according to an exponent ν below or equal to one. The cost share of servers and other compute equipment provides the calibration intuition: doubling spending does not necessarily double productive capacity, because construction, power, cooling, and deployment constraints intervene.
The resulting structural comparisons are:
Compute ratio = λ<sup>ν/(1−νβ)</sup>
Investment ratio = λ<sup>1/(1−νβ)</sup>
The exact expressions are less important than their relationship. The dollar investment ratio is larger because spending must overcome the concavity of producing capacity as well as the contest wedge. A private race can therefore create a moderate gap in effective compute and a larger gap in the capital committed to produce it.
At the BIS paper’s baseline, there are five stylised frontier coalitions, β is set to 0.80, and ν to 0.70. Those values produce a contest wedge of 1.20. In the frictionless benchmark that isolates the contest, compute is about 1.34 times the planner level and dollar investment is about 1.51 times the planner level. The figures are not measurements of individual firms’ efficiency. They are the output of a structural model under assumptions about the number of coalitions, demand curvature, and production concavity.
The diagram separates three effects that are often compressed into the phrase “AI investment.” Total compute can increase sector revenue. A coalition can improve its position relative to rivals. And the cost of producing capacity can rise faster than capacity itself. The first effect explains why investment has value. The second explains why private incentives can be too strong. The third explains why the spending gap can exceed the capacity gap.
What the baseline does—and does not—establish
The full calibrated model is not identical to the frictionless comparison. It adds the financing and timing choices that later shape the bust. Even so, its period-one equilibrium commits about $280 billion per coalition, or approximately $1.4 trillion across five coalitions. The planner benchmark in the same environment commits about $200 billion per coalition, or roughly $1.0 trillion across the sector. The full-model ratio is therefore about 1.40, close to the structural 1.51-times dollar comparison once the additional choices are included.
These numbers are useful for scale, not for precision forecasting. They show that a contest wedge of 1.20 at the margin can become a materially larger difference in spending when production is concave. They do not show that the actual AI sector is exactly 40% above an observable social optimum. The model uses a stylised coalition count, an assumed demand curvature, and a productivity scale calibrated to the observed capex environment. The authors deliberately choose β = 0.80 as a conservative upper bound relative to their discussion of usage and revenue growth; lower values of β produce larger distortions, with the ratio approaching three in the paper’s central lower-curvature readings.
The interpretation should therefore be asymmetric. Evidence of rapid AI use makes a large build-out plausible. It does not remove the contest wedge. Conversely, the presence of a contest wedge does not prove that the build-out will fail. A high-productivity outcome could generate enough revenue to cover a large capacity expansion. The model says that the private race creates a tendency to commit more resources than the sector would choose if it internalised the transfer between rivals and the decline in marginal sector returns.
This is also why the question cannot be settled by asking whether the leading firms are individually competent or whether their current products have customers. Competence and demand can coexist with excessive rivalry. A firm may need to build aggressively to avoid losing a strategic position even when it recognises that the sector as a whole cannot earn the same return on every new unit of capacity. Private rationality does not guarantee collective stability when the payoff includes relative position.
Scale is only the first dimension of the problem. The same contest can influence how firms finance the capacity and when they commit it. An investment made with internal cash after demand is known has a different exposure from an investment pulled forward with debt, a purchase commitment, and an equity tie to a partner. The next question is therefore not simply how much the race builds, but how the race finances and times what it builds.