Picks, Shovels, Superstars, and Poachers: Explaining the AI Investment Boom

Standard real-options theory makes a clear prediction: when technology advances rapidly, costs fall steeply, and technical approaches remain uncertain, firms should wait. The option value of deferring irreversible investments dominates the value of moving early. We observe this pattern across industries with rapid obsolescence—semiconductor firms stage their fab investments, biotech companies defer scaling until Phase II data, and hardware manufacturers maintain flexible supply chains.

Yet the current AI landscape defies this logic. Instead of cautious experimentation and staged investment, we observe firms committing billions to training runs, engaging in fierce bidding wars for talent, and locking in massive compute contracts. OpenAI, Anthropic, Google, Meta, and others are simultaneously pursuing similar technical approaches with extraordinary capital intensity.

This article develops a model to resolve this puzzle. The key insight is that overoptimism about market size, the structure of winners, combined with scarce inputs and slow belief updating, can generate a race equilibrium where early resource holders—infrastructure providers, top engineers, and intermediaries like recruiters—capture rents while competing firms dissipate value.

Model Setup

Consider N firms competing for a transformative market opportunity with true value V. Each firm i perceives this opportunity as $\hat{V}_i(t) = \alpha_i(t) \cdot V$ where $\alpha_i(t) \geq 1$ represents overoptimism relative to fundamental value.

Belief Dynamics and Learning Impediments

Beliefs evolve according to: $$\dot{\alpha}_i(t) = -\lambda(\alpha_i(t) - 1)$$

where $\lambda$ represents the base learning rate. However, the AI domain presents unique challenges for belief updating:

1. Delayed payoff structure: True commercial value may not materialize for years
2. Noisy intermediate signals: Capability improvements don't directly translate to economic value
3. Multiple interpretations: The same evidence supports both optimistic and pessimistic narratives

We capture this through an effective learning rate $\lambda_{\text{eff}} = \lambda(1-\eta)$ where $\eta \in [0,1)$ represents the strength of delayed-payoff beliefs. As $\eta \to 1$, learning effectively stops despite continued evidence accumulation.

Production Technology and Input Markets

Firms combine two categories of inputs to generate effective capability:

$$E_i = a_s S_i + a_o O_i$$

where:

• $S_i$ represents specialized human capital (ML researchers, engineers with specific domain expertise)
• $O_i$ represents scalable infrastructure (compute, data centers, training capacity)

The Talent Market

The supply of capable ML engineers (or ML engineers thought to be capable) is essentially fixed at $\bar{S}$ in the short to medium term.

The market for this talent exhibits several key features:

Training Investment Deterrence: When firm j invests $T$ to develop an engineer's capabilities, generating productivity value $R$, the investment's expected value depends critically on retention probability. With poaching probability $q$, the training NPV becomes:

$$\text{NPV}_{\text{train}} = (1-q)R - T$$

As market competition intensifies, $q$ rises, potentially making $\text{NPV}_{\text{train}} < 0$ even when $R > T$. This generates a classic hold-up problem: socially valuable training doesn't occur because firms cannot capture the returns.

Bidding Equilibrium: Instead of expanding the talent pool, firms bid for the fixed supply. In equilibrium, wages approximately equal the marginal contribution to contest success:

$$w^* \approx \frac{\partial P_i}{\partial E_i} \cdot a_s \cdot \hat{V}_i$$

Overoptimism ($\alpha_i > 1$) directly inflates wage offers, creating a transfer from firms to talent.

Infrastructure Markets

The infrastructure input market exhibits upward-sloping supply:

$$p = \phi_0 + \phi_1 \sum_i I_i$$

This reflects both physical constraints (fab capacity, energy availability) and market power among suppliers. Unlike talent, infrastructure can expand over time, but with significant lags and increasing marginal costs.

Contest Structure and Equilibrium

Firms compete in a Tullock contest where success probability follows:

$$P_i = \frac{E_i^r}{\sum_j E_j^r}$$

The parameter $r > 0$ governs how effort differences translate to success probabilities. Each firm maximizes:

$$\Pi_i = P_i \cdot \hat{V}_i(t) - w S_i - p \cdot I(O_i) - c \cdot O_i$$

The first-order conditions reveal that overoptimism acts as a multiplier on input demand. Firms aren't irrational—they're optimizing given their beliefs—but collective overoptimism drives input prices above fundamental values.

Rent Distribution During the Boom

Three groups capture rents during the competition phase:

1. Specialized Talent

Engineers with scarce expertise earn $w^*$, which increases in:

• Degree of overoptimism ($\alpha$)
• Perceived market value ($V$)
• Talent scarcity ($1/\bar{S}$)

These payments occur immediately, in cash, regardless of eventual technical success.

2. Infrastructure Providers

Vendors capture producer surplus: $$\pi_{\text{infra}} = \int_0^{Q^*} (p(q) - MC(q))dq$$

where $Q^* = \sum_i I_i^*$. Market power and capacity constraints allow sustained margins above competitive levels.

3. Market Intermediaries

Recruiters and search firms facilitate the reallocation of scarce talent, earning fees proportional to placement compensation: $$\pi_{\text{recruit}} = \rho \cdot w^* \cdot M$$

where $\rho$ represents fee rates (typically 0.2-0.3) and $M$ represents transaction volume.

Importantly, recruiters provide genuine value by improving match quality and reducing search frictions. Their rents reflect the value of intermediation in a market with extreme information asymmetries and urgent talent needs. The social inefficiency arises not from their participation but from the excessive churn driven by the poaching equilibrium.

Boom Duration and Belief Convergence

Starting from initial overoptimism $\alpha_0$, the belief trajectory follows:

$$\alpha(t) = 1 + (\alpha_0 - 1)e^{-\lambda(1-\eta)t}$$

The time to approach fundamental values (reaching $\alpha \approx 1 + \epsilon$) is:

$$T_{\text{boom}} \approx \frac{1}{\lambda(1-\eta)} \ln\left(\frac{\alpha_0 - 1}{\epsilon}\right)$$

This expression reveals three key drivers of boom persistence:

1. Initial overoptimism magnitude ($\alpha_0$): Greater initial expectations take longer to correct
2. Learning rate ($\lambda$): Noisier signals and multiple interpretations slow convergence
3. Delayed payoff belief ($\eta$): Strong narratives about future breakthroughs can sustain booms indefinitely

When $\eta$ approaches 1, the denominator approaches zero, and the boom duration extends dramatically—even with rational Bayesian updating.

The model distinguishes between transfers and deadweight losses:

Transfers

• Wage payments to engineers: $w^* \cdot \bar{S}$
• Infrastructure rents: $(p^* - MC) \cdot Q^*$
• Intermediary fees: $\rho \cdot w^* \cdot M$

These represent redistributions from firms to input suppliers. While they affect the distribution of surplus, they don't necessarily reduce total welfare.

Deadweight Losses

1. Duplicative research efforts: $\sum_i c \cdot O_i$ where much of this represents parallel, secret development
2. Misallocation of general-purpose compute: High-value scientific computing crowded out by speculative AI training
3. Underinvestment in human capital development: $(R - T) \cdot N_{\text{potential}}$ in forgone training
4. Transition costs from excessive churn: Productivity losses during job switches, estimated at $\tau \cdot M$ where $\tau$ represents per-transition cost

The key insight is that overoptimism amplifies these deadweight losses by increasing both their magnitude and duration.

Resolving the Initial Puzzle

Why do we observe overinvestment rather than the option-value-driven underinvestment predicted by standard theory?

The model identifies four key factors:

1. Winner-take-all beliefs: Unlike marginal technological improvements, firms believe AI represents a categorical shift where the leader captures enormous value. The perceived cost of waiting exceeds the option value.
2. Strategic complementarity in the talent market: Once some firms begin bidding for talent, others must follow or risk being locked out entirely. The individually rational response is to bid aggressively, even knowing collective overinvestment results.
3. Slow learning with maintained optimism: The combination of noisy signals and strong priors means firms can maintain optimistic beliefs for extended periods, sustaining investment despite mounting costs.
4. Inelastic input supply: Unlike industries where investment can create new capacity, the binding constraints (top talent, near-term compute) cannot be quickly expanded. This transforms the problem from an investment decision to a bidding contest.

Conclusion

This model explains the apparent paradox of massive, simultaneous investment in an uncertain technology with rapid obsolescence. The combination of overoptimism about market size, scarce inputs that cannot be quickly expanded, and impediments to learning creates conditions where racing equilibria dominate waiting equilibria.

The primary beneficiaries are not the competing firms but those who supply scarce inputs—infrastructure providers, specialized engineers, and the intermediaries who facilitate resource reallocation. These groups capture rents during the boom phase, paid in certain cash flows, while firms bear the risk of eventual market correction.

The framework suggests that standard policy prescriptions—such as subsidizing R&D or reducing barriers to entry—may actually amplify inefficiencies by adding more contestants to an already over-dissipated prize. Instead, interventions that expand the supply of binding constraints (training programs for ML engineers, increased compute capacity) or improve information quality (standardized benchmarks, required disclosure) would more effectively reduce deadweight losses.