"Geopolitics in the Age of Artificial Intelligence" is a Foreign Affairs article published on 2026-01-27 by Jake Sullivan and Tal Feldman. It argues that US AI strategy should be designed for uncertainty rather than built on a single prediction about how the technology will develop, and proposes a 2x2x2 scenario matrix, the Eight Worlds Framework, to test strategy against multiple possible futures. Sullivan served as National Security Adviser from 2021 to 2025, and the article reflects his post-government thinking.
Summary of argument
Sullivan and Feldman argue that every AI policy debate rests on hidden assumptions about how the technology will develop, and that Washington should stop betting on a single prediction and instead adopt a framework for decision-making under uncertainty. The proposed framework, the Eight Worlds Framework, makes these assumptions explicit by combining three binary axes into eight possible "worlds," each implying a different US posture.
The three axes are:
- Nature of AI progress: whether AI reaches superintelligence or plateaus at bounded and jagged intelligence.
- Ease of catching up, the Fast-Follow Problem: whether breakthroughs can be quickly copied, or whether frontier capability depends on a full stack that is hard to reproduce.
- China's strategy: whether Beijing is racing aggressively for the frontier or prioritizing deployment and commoditization of US breakthroughs.
Each axis has two plausible answers, yielding eight possible worlds.
The Eight Worlds
| World | Superintelligence? | Catch-up? | China racing? | US posture |
|---|---|---|---|---|
| 1 | Yes | Hard | Yes | Arms race / Manhattan Project 2.0 |
| 2 | Yes | Hard | No | Unipolar AI moment — lead responsibly |
| 3 | Yes | Easy | Yes | All-out proliferation — race defensively + resilience |
| 4 | Yes | Easy | No | Fleeting unipolar window — act fast |
| 5 | No | Hard | Yes | Grinding innovation race + diffusion |
| 6 | No | Hard | No | Comfortable US lead — focus on prosperity and safety |
| 7 | No | Easy | Yes | Diffusion race — spread systems before China does |
| 8 | No | Easy | No | Technology contest resembles 5G — deployment and scale |
Key claims
The authors hold that compute is foundational: control over chips, data centers, and energy determines who trains and deploys frontier systems (Source: Compute Governance). They argue that AI Diffusion is as strategically important as frontier capability, on the reasoning that the systems that take root globally will determine whose values define the digital order. They treat Export Controls (AI) as context-dependent, useful when catching up is hard but less effective or even counterproductive when copying is easy.
Sullivan and Feldman also argue that private sector incentives diverge from national interest, in that labs bet on superintelligence and prefer overseas infrastructure, and that managing this tension is a core government task. They frame risk management as a source of legitimacy rather than only a constraint, on the grounds that it prevents competition from collapsing due to accidents, misuse, or loss of control. They contend that allies multiply US capacity, improving the odds that democratic systems define the AI age.
Sources of US AI power
The article identifies six sources of US AI power: compute (chips, data centers, energy); robotics and advanced manufacturing, which translate digital intelligence into physical capacity; the industrial-scientific base of R&D, talent, manufacturing, and energy; risk management, encompassing safety, accountability, and political sustainability; diffusion, meaning the embedding of US systems globally before Chinese alternatives take hold; and allies and partners, which multiply capacity through coordination.
Implicit policy tools
The article points to a set of policy tools, several left implicit. These include export controls and investment restrictions, which act as an implicit subsidy to US industry; government signaling and expectation-setting; R&D tax credits, infrastructure investment, and federal grants; immigration policy as a means of talent attraction; federal procurement as a demand signal; the Development Finance Corporation for overseas AI deployment; and decisions between open-source and proprietary models.
The authors note that policymakers' own decisions shape which future emerges, especially through export controls. They acknowledge that China may not have a single coherent national plan, and they describe the framework as designed to evolve, with new axes able to replace resolved questions.