Decisive Economic Advantage
Why do technological leads usually dissipate and what would have to be true for AGI to break this pattern?
History’s commanding technological leads have always eroded. Britain’s machinery monopoly and America’s semiconductor dominance both fell to diffusion, imitation, and diminishing returns.
DeepSeek-R1’s January 2025 release seemed to confirm the same pattern for AI. A Chinese lab, operating under U.S. export controls, appeared to match frontier American models not by matching compute but through efficiency gains, at a fraction of the cost. For many, convergence feels inevitable.
But convergence in frontier capabilities and convergence in economic outcomes are different things. Our report Decisive Economic Advantage asks whether an initial lead could compound rather than erode, and through what mechanism.
The answer shapes how we should interpret events like the DeepSeek moment, which is a clue to whether the race equilibrates or that one particular pathway to dominance was temporarily blocked.
From DSA to DEA
The concept grows out of Nick Bostrom’s Decisive Strategic Advantage (DSA). A DSA refers to the idea that an AI system, or the actor controlling it, could achieve capabilities so far beyond any competitor that it could prevent others from catching up and permanently shape the future according to its own values.
A Decisive Economic Advantage (DEA) is an economic analog, representing a competitive regime in which feedback mechanisms cause an initial lead to compound over time, progressively constraining a follower’s ability to restore balance.
This is not the same as being ahead. The question is whether the competitive system is converging (the lead shrinks over time) or diverging (it grows without bound). A DEA marks the point where dynamics have shifted irreversibly toward the latter.
Economic dynamics operate on different timescales than capability dynamics, involve different actors, and create different intervention points, so even if frontier AI capabilities converge, economic feedback loops might not.
The DeepSeek story can be instructive here as well. Efficiency gains that close capability gaps do not necessarily close deployment gaps, and they certainly do not close infrastructure gaps. A world where every country can build frontier models is not necessarily a world where competitive dynamics equilibrate.
Multiple Pathways to Decisive Advantage
In the report, I model competition between a leader and a follower across three dimensions (AI capabilities, economic deployment, and compute infrastructure) and identify three distinct mechanisms through which a decisive advantage can emerge.
Intelligence explosion. If better AI accelerates AI progress, a leader’s advantage compounds exponentially. The model confirms this produces runaway divergence. This is the scenario many have in mind when they worry about decisive advantage, but it turns out to be only one of several pathways.
The development flywheel. Large-scale deployment generates real-world data that improves models, which enables further deployment. No recursive self-improvement required, just deployment experience feeding back into technology.
The hardware moat. AI income advantages get reinvested into compute infrastructure, enabling faster model improvements and greater deployment, generating more income. Hardware constraints become moats rather than bottlenecks. One interesting aspect about this pathway is that it can be self-reinforcing even if the underlying technology diffuses relatively freely, because the economic gap funds infrastructure accumulation faster than rivals can respond.
These pathways have different timelines and respond differently to policy.
What the Simulations Show
I ran 4,000 parameter combinations under deep uncertainty and defined a DEA quantitatively as reaching a 100x income gap with the system still diverging (DEA-100). Three findings stand out.
Decisive dynamics do not require exceptional assumptions. Between 40% and 60% of simulations reached the DEA-100 threshold, depending on initial advantage duration and capital adjustment regime.
Frontier capabilities are the most likely, but not the only, pathway to DEA. Among DEA-100 trajectories, 96% involved frontier self-reinforcement dynamics. But 82 of 4,000 trajectories reached permanent economic dominance through accumulation alone, overwhelmingly through the hardware moat dynamic. Pure development flywheels were vanishingly rare, which makes intuitive sense as they require a very narrow set of conditions.
There is a viable policy window. The median time to DEA-100 via intelligence explosion is about 10 years, though some happen as quickly as three years; via accumulation, about 14.5 years.
The next question is whether actors can correctly identify the regime before the window closes.
The Window Closes Fast
I tested two stylized follower responses: “full stack denial” (slowing the leader’s frontier progress and choking scaling capacity) and “ecosystem containment” (limiting diffusion and weakening the development flywheel). I triggered these policies in the model only once the income gap becomes apparent to the follower.
The consistent finding across both policies and mechanisms was that earlier, weaker interventions are more effective than stronger interventions enacted later. The structural reason is because feedback loops compound, so early action operates on a smaller gap with weaker reinforcing dynamics. A policy that works at 2x income gap may fail at 5x or 10x.
Pathway also matters. Full stack denial avoided decisive outcomes in over 90% of accumulation-driven cases even at a 5x gap, but in only 12% of intelligence explosion cases. Correctly diagnosing which mechanism is operating is as important as acting early.
One important scope caveat: the model traces income gaps, not military or geopolitical power. It is a two-bloc model that does not capture third-party actors, alliances, or broader institutional responses. The results characterize structural dynamics rather than calibrated forecasts.
Why This Matters
Discussions of transformative AI tend to focus on capabilities, which makes sense given that monitoring capabilities is a critical first step to understanding when transformative potential emerges. But economic dynamics can shape geopolitical competition even when frontier capabilities are converging.
A world where the intelligence explosion doesn’t happen is not necessarily a world where competition equilibrates.
The strategic implications cut both ways. For a leader, decisive advantage may not require winning the frontier race. Winning the deployment and infrastructure races may suffice. For a follower, the response window is narrow and the right intervention depends on which mechanism is operating.
China’s demonstrated capacity for large-scale deployment and infrastructure buildout offers pathways to an advantage that doesn’t depend on frontier models. Export controls designed to deny frontier capabilities may therefore be necessary but not sufficient. If the hardware moat mechanism matters more than the intelligence explosion pathway, the relevant competition is over deployment scale and investment capacity, not just who reaches the frontier first.
DeepSeek demonstrated that the capability gap could be contestable. What remains uncertain is whether it remains contestable in the future, and whether the deployment gap and infrastructure gap are contestable.
Britain’s industrial lead diffused because the underlying dynamics favored convergence despite active efforts to prevent it. Whether AI follows the same pattern depends on which feedback mechanisms dominate. By the time we’re certain of the answer, it may be too late to shape the outcome.