Why AI labs race: a model separates leadership prizes from shared risk

A company can benefit from developing advanced AI sooner even when the industry would collectively prefer a slower pace. In Prizes, Shared Hazard, and the Speed of Frontier AI, Texas A&M economist Li Gan models two reasons: competition for commercial leadership and an incentive problem around shared catastrophe risk.

The November 2026 research is an arXiv preprint, not a peer-reviewed finding. Its numerical results describe illustrative scenarios rather than measured company behavior or estimated probabilities of catastrophe.

Two reasons to accelerate

Each modeled lab assigns a constant share of computing capacity to capability research. Safety research receives a fixed share, and customers receive the remainder. More capable AI helps accelerate further capability research, while safety knowledge accumulates over time and reduces danger.

Slowing development therefore has competing effects. A lab reaches later capabilities with more safety knowledge, but spends longer exposed to the risks of existing systems. Delay is not automatically beneficial: the paper gives cases where additional exposure outweighs the safety benefit.

The commercial incentive comes from a prize whose allocation depends on relative arrival dates. An earlier arrival improves a lab’s chance of winning, although it does not guarantee victory. If every company advances equally quickly, nobody gains a relative advantage.

The second force is what Gan calls the hazard externality. A catastrophe originating at any frontier destroys every lab’s terminal stake. Yet an individual lab balances its own timing and competitive gains against only the marginal hazard contribution assigned to its frontier. A coordinated industry instead weighs changes across all frontiers together.

In the baseline, industry hazard at a common research share is held constant as the number of labs changes. That normalization isolates strategic incentives; it is not an estimate of the physical risk created by adding another real laboratory.

What the headline numbers mean

The baseline assumes three labs, an 18-month full-speed takeoff period and a 75% catastrophe probability during that period. The takeoff runs from AI that automates machine-learning engineering to AI that surpasses human scientists across domains. These are scenario inputs, not predictions established by the paper.

Gan also imposes a racing allocation of 50% of compute to capability research, then fits the contest’s sensitivity to make that allocation an equilibrium. Consequently, the model does not independently discover that companies will choose this speed.

Under common-share coordination, the same calibration produces a takeoff lasting about 69 months and a catastrophe probability of about 6%. Removing the speed-dependent commercial contest, while retaining the hazard externality, produces roughly 55 months and 12% risk.

Those comparisons suggest that neither force alone explains the imposed racing allocation. Together, they reinforce the incentive to accelerate.

But attributing the risk gap requires care. Removing the contest first closes about 91% of the baseline gap between racing and coordination. Removing it after eliminating the hazard externality assigns it only 39%. Averaging the two removal orders gives 65%.

The conclusion also changes with assumed risk. In a separately refitted scenario with 25% full-speed catastrophe probability, the contest receives 69% of the gap when removed first, but 44% averaged across both orders. There is no universal finding that commercial rivalry explains almost all AI risk.

Common rules versus individual restraint

One cautious lab achieves much less than an industry-wide rule in the baseline. When it adopts the coordinated research share, its rivals reduce theirs from 50% to about 35%. Industry catastrophe probability falls to roughly 57%, delivering only 27% of the risk reduction achieved when every lab adopts that same restriction.

An idealized sharing arrangement also aligns incentives with coordination. Every capability finding must reach every lab immediately, creating one shared frontier with no private head start. The result concerns aggregate research input, not a uniquely determined contribution from each company. Delayed releases, private research and access by foreign rivals fall outside that ideal.

Gan applies the model to the September 2026 White House Accord described in the paper, arguing that safety audits leave competitive speed incentives intact when research shares remain unrestricted. The model does not evaluate the effectiveness of monitoring or oversight itself.

A benchmark, not a policy verdict

Coordination here maximizes the labs’ joint terminal payoff, not social welfare. Consumer surplus, service revenue and catastrophe losses outside those stakes are excluded. Research shares are stationary, compute grows exogenously, and the model counts danger only during the takeoff period.

These restrictions matter. The paper offers a framework for distinguishing incentives that safety requirements and pacing rules address—not proof that any particular agreement is enforceable, socially optimal or sufficient to make advanced AI safe.