How Experts Think AI Will Change the Global Balance of Power by 2035

The futures of artificial intelligence (AI) and global order are both contested, as is the relationship between them. Policymakers face the challenge of navigating a future defined by profound technological uncertainty. To help them do so, this project examines how the concentration or diffusion of the capability to produce “frontier AI”—AI that pushes the leading edge of what current systems can do—and the governance structures that emerge around it will shape the distribution of power and the rules-based international order by 2035.

Below are plausible alternative futures that government policymakers, businesses, and civic leaders can use to strategize effectively.

To map the landscape of expert opinions on the future of AI and global order, CFR surveyed foreign policy leaders with knowledge of AI, security studies, governance, civil society, and law, representing more than three hundred fifty voices from around the world. Experts were asked to reason about frontier AI capabilities, which are a moving target.  

Fragmentation and Concentration Projections

The survey asked respondents to predict how concentrated AI capabilities would be and how coherent AI governance would be in 2035 by using sliders to offer a score from zero to one hundred.

Governance Projections

There was broad consensus among experts that AI governance is unlikely to cohere. More than 80 percent of experts ranked their expectation of governance coherence as low (below fifty on a zero-to-one hundred scale), and 52 percent specifically offered a score below thirty. Responses are clustered at the lower coherence end in a broad range near 30 (mean of 32, median of 29), with a modest tail toward higher coherence.

Capability Projections

There was no clear consensus among surveyed experts with respect to the distribution of frontier AI capabilities. Many (46 percent) foresee continued concentration of AI capabilities, defined as two to three actors capable of developing frontier models, recognizing that frontier is a moving target. Conversely, more than half (54 percent) foresee the diffusion of frontier capabilities to dozens of states and nonstate actors. 

Experts were opinionated: responses are bimodal, clustering separately toward each end of the scale, with the mid-range least concentrated. Only 10 percent clustered near the midpoint (offering a score between forty and sixty on a one-hundred-point scale). That could reflect disagreement over whether scaling laws will hold, or different understandings of what constitutes diffusion of frontier AI.

Confidence Level

Respondents expressed a high degree of epistemic humility. That likely reflects the difficulty inherent in making predictions over the span of a decade, especially about rapidly evolving, novel technologies.

Events’ Impact on Capabilities

The survey asked respondents to consider what could change their estimates of how diffused or concentrated frontier AI capabilities will be by 2035.

Unsurprisingly, more than 60 percent said that the nationalization of a frontier AI lab or serious accident with national security or public safety consequences would lead to a concentration of capabilities. More than 65 percent said a major open-source frontier model release or the emergence of a Global South compute coalition would lead frontier AI capabilities to diffuse.

Notably, more than half (53 percent) of respondents indicated that a U.S.-China conflict would lead to greater concentration of AI capabilities, perhaps because it would cause Washington and Beijing to deepen their AI investments and more tightly guard their AI advancements. This response was somewhat unexpected, given that Washington and Beijing are already competing fiercely.

Surprisingly, 43 percent indicated that a binding international treaty would lead to a concentration of AI capabilities. Respondents could have used nuclear analogies in their reasoning, assuming the treaty would be like an arms control agreement from the Cold War era—designed to limit the number of actors and the proliferation of frontier AI labs. Respondents were more evenly split over whether large-scale protests over AI would lead to concentration (22 percent) or diffusion (25 percent), with 54 percent selecting they would have minimal effect and only 5 percent each selecting a strong shift toward diffusion or concentration.

Events’ Impact on Governance

Respondents were asked to consider what could change their estimates of how coherent or fragmented AI governance will be by 2035.

More than 70 percent of respondents believe that a binding international treaty or serious AI accident would drive toward more coherent governance. A plurality (44 percent) foresees anti-AI protests having the same effect. However, many other experts were pessimistic about the effect demonstrations would have, with 35 percent expecting anti-AI protests to have minimal impact on governance coherence. Only 9 percent expect an AI accident to have minimal consequences.

More than half (53 percent) believe a U.S.-China conflict would fragment governance. Any outright conflict is likely to accelerate AI development and deployment and put a U.S.-China bilateral agreement—which respondents see as the most likely mechanism for achieving a broad and binding governance regime—out of reach.

The survey asked respondents a series of corresponding policy questions about AI diffusion and governance, including who will have the power to implement technology standards, how the benefits of AI will be distributed, and what AI’s impact on strategic stability will be, among others.

Governance Framework Driven by Bilateral Agreement

Reflecting the technological and political capabilities held by Washington and Beijing, a plurality of respondents thought the most likely catalyst for a binding multilateral agreement on AI governance would be a U.S.-China bilateral agreement. The European Union, which already has robust regulatory frameworks for data privacy, cybersecurity, content moderation, and digital competition, was also expected to be an important driver of global governance. 

Actors That Set Standards

Respondents expect the fragmentation of global governance to be mirrored at the operating level, as different markets and governments adopt competing standards of accountability, security, transparency, and safety for frontier AI models.

Nonstate Leverage

When heads of state met for the annual Group of Seven meeting in June 2026 in the French Alps, they were joined by the CEOs of OpenAI, DeepMind, Anthropic, the French AI company Mistral, and other AI companies. Survey respondents would not be surprised that those CEOs attended the meeting, as nearly 70 percent believe that frontier labs will be the nonstate actors that gain the most political leverage from AI.

The Distribution of Productivity Gains

Policymakers in many countries are discussing the need for sovereign AI to ensure insulation from outside pressure and the widespread distribution of AI benefits. Survey respondents are skeptical that efforts in the Global South will succeed, with a little more than 50 percent expecting the economic gains to be concentrated in the hands of a few in the most developed economies. Respondents see white-collar work, including information technology and back-office services, as the sector most likely to suffer. They were more optimistic about AI’s effects on scientific discovery and health care.

Global South Development Pathways

If the Global South is to emerge as a major player in AI, most respondents do not believe it will be the result of new alliances or smart political maneuvering. Instead, it will be because of a major shift in the trajectory of frontier AI development, from models with heavy data, energy, compute, and capital requirements to smaller, less energy-demanding edge ones.

Military Systems and Great Power Conflict

Respondents believe the loss of control of autonomous systems, accelerated cyberattacks, rushed decision-making that causes mistakes and miscalculations, and other risks of integrating AI into military systems make great power conflict more likely.

The Domestic Institutional Adaptation Gap

In a March 2026 survey, 74 percent of Americans thought the government was not doing enough to regulate the use of AI. CFR survey respondents were similarly concerned, with only 5 percent believing that domestic institutions were keeping pace with the development of AI. Job displacement, democratic backsliding, and regulatory capture by powerful AI firms are top concerns.

This survey was conducted from April 30, 2026, to June 16, 2026. The results reflect 362 responses in total, 237 of which addressed all sections and 119 of which answered only the first 2.

Respondents were recruited via email outreach to individual and corporate members of CFR and members of the Council of Councils, a CFR initiative connecting leading foreign policy institutes from around the world. That survey pool ensured the results captured expert expectations, rather than a statistically representative sample of the broader population.

Professional Domain and Sector

Of the respondents who filled out optional demographic information (232), the largest group listed their primary professional domain as international relations or security studies (28 percent), followed by AI or computer science research (17 percent), “other” (15 percent), government or multilateral organization (13 percent), economics or development (10 percent), governance or international law (9 percent), and civil society or advocacy (8 percent).

In terms of sectors, 33 percent listed their primary sector as the private sector (general), 22 percent as academia or research institution, 12 percent as government (national), 11 percent as think tank or policy institute, 10 percent as private sector in the AI lab or infrastructure fields specifically, 6 percent as civil society or nongovernmental organization, 4 percent as multilateral or international organization, and 2 percent as “other.”

Primary Geographic Base

The majority of respondents (84 percent) were North America–based, while 5 percent listed Europe, 6 percent listed other or global, 3 percent listed East Asia and the Pacific, and under 1 percent each are based in Sub-Saharan Africa, South and Southeast Asia, Middle East and North Africa, or Latin America and the Caribbean.

Age and Gender

In terms of age, 41 percent of those who answered (231) are sixty-five years or older, 28 percent are thirty-five to forty-nine, 28 percent are fifty to sixty-four, and 1 percent are under thirty-five. In terms of gender, 68 percent are male, 29 percent are female, and 3 percent preferred not to say.

Younger experts were more likely to expect a greater concentration of AI capability. Of sixty-five respondents aged thirty-five to forty-nine, 61 percent expect concentration over diffusion with a median of sixty-seven (scoring concentration above fifty on the scale), and of sixty-four respondents aged fifty to sixty-four, 53 percent expect concentration. Meanwhile, only 32 percent of ninety-five experts aged sixty-five and older expect concentration of advanced AI capabilities. This is one characteristic of the split in capability expectations.

Familiarity With AI and International Relations

The survey asked respondents to assess their familiarity with AI technical developments and international relations or global governance on a scale from one (limited) to five (deep expertise).

Of the respondents, 39 percent assessed themselves as familiar (four) or deeply familiar (five) with AI technical developments. The highest number, 34 percent, rated themselves as moderately expert (three), with 28 percent assessing themselves at four, 19 percent at two, 11 percent at five, and 8 percent at one.

A much higher percentage, 73 percent, ranked themselves as familiar (four) or deeply familiar (five) with international relations or global governance. The largest percentage, 37 percent, rated themselves at four, followed by 36 percent at five, 20 percent at three, 6 percent at two, and 0.9 percent at one (limited).

LLM Usage

Respondents were asked about their use of large language models (LLMs) in their professional and personal lives (daily, several times a week, occasionally, or rarely or never). In professional life, 55 percent use LLMs daily, 23 percent use them several times a week, 12 percent use them occasionally, and 10 percent use them rarely or never. In personal life, 46 percent use LLMs daily, 20 percent use them several times a week, 19 percent use them occasionally, and 14 percent use them rarely or never.

Answer Coding

In the first section of the survey, respondents were asked to answer a question about the concentration or distribution of frontier AI by moving a slider on a scale from zero to one hundred, with highly concentrated (two to three actors capable of building frontier AI systems) at zero and widely distributed (dozens of actors with independent frontier AI development capacity) at one hundred. The second question involved ranking AI governance from fully coherent (all major countries adhere to common rules with enforcement mechanisms) at zero to fully fragmented (each major country operates its own competing framework) at one hundred. Charts are oriented for readability with “concentrated” and “coherent,” as well as “diffused” and “fragmented,” labeled at the stated ends.

Further Information

Download the survey questions.

Download the survey data.

Subscribe to receive updates from this project on AI and the next global order in your inbox.