Who Will Win the Race to AGI or End Up Self-Destructing?

In the tech world, real options theory says that when uncertainty is sky‑high, it’s wiser to wait before investing. Yet, in the race to build artificial general intelligence (AGI), the US and China are speeding up funding, even though a mistake could be catastrophic. How can this be explained? A new study models the AGI race as a continuous‑time preemption game where global ruin risk is tied to development speed. Here, the cost of failure isn’t just sunk investment but a systemic ruin parameter (D) that grows with rapid research and affects everyone. Because both players share the fear of disaster, the risk term cancels out in the equilibrium condition, creating a “suicide region” where competitive pressure forces early AGI deployment despite a negative risk‑adjusted net present value. Small incidents, or “warning shots,” fail to halt the race because the prize is winner‑takes‑all. The only way to reverse this trend is to internalize the ruin cost, making safety research an economic prerequisite. The authors suggest private liability thresholds and design mechanisms to restore the option value of waiting and ensure safer AGI development.

The proposed model employs a continuous‑time preemption game formulation, where two sovereign actors (USA and China) choose the activation time τ of AGI. Each player's payoff is V_i(τ)=e^{-rτ}(B_i - I) - D·f(τ), with r discount rate, B_i expected benefits, I investment cost, and D systemic damage. The function f(τ) captures catastrophe probability, increasing with development speed. In Nash equilibrium, players are indifferent between waiting or investing, leading to ∂V_i/∂τ=0. Because the term D·f(τ) appears in both payoffs, the risk component cancels, leaving an indifference condition identical to a risk‑free game. This generates the “suicide region”: an interval of τ where, despite a negative net expected value, competitive pressure forces early activation. Moreover, the model shows that warning signals (minor disasters) do not shift the τ threshold, as the winner‑takes‑all structure keeps the marginal payoff derivative unchanged. To break the mechanism, a internalized ruin cost must be introduced, e.g., via a private liability tax D̂ ≥ D_c, where D_c is the critical threshold solved by setting V_i(τ)=0 with D̂ inserted. This intervention restores a positive option‑to‑wait value, incentivising actors to invest first in safety research before deploying AGI. The paper also proposes design mechanisms such as global insurance contracts and multilateral governance rules to ensure safety is a prerequisite for commercial AGI implementation.