The AI Balance of Power
The influence of artificial intelligence (AI) on geopolitics, though likely great, will depend on the technology’s trajectory. Many leaders in the frontier labs believe that scaling laws—the concept that model performances increase at predictable rates as model size, datasets, and computational power increase—will continue to hold. If they do, then superintelligence, a hypothetical AI that surpasses human intelligence across all fields, could be reached in the next three-to-five years. That would transform geopolitics, U.S. strategy, and the entire strategic landscape. But superintelligence is not a given, and other AI researchers are more skeptical.
In any case, estimating how and when new capabilities will emerge is difficult. Technological progress will require different types of models and new breakthroughs, as well as dealing with unexpected bottlenecks that could slow progress. In addition, the capabilities to develop new models, as well insight about their strengths and weaknesses, reside primarily in the frontier labs, making them both indispensable partners for states and powerful geopolitical actors in their own right. Any successful U.S. strategy will need to prepare for an uncertain distribution of AI power among state and nonstate actors.
The AI world is currently bipolar, with the United States and China dominant. In 2024, the United States had forty notable models, China fifteen, and Europe three. GPU cluster performance is a measure of how effectively graphics processing units, circuits that are used to train and deploy AI models, work together. As of May 2025, the United States had about 75 percent of global GPU cluster performance, with China in second place with 15 percent. The United States attracts the world’s most talented AI researchers, whereas China trains close to 50 percent of the world’s top-tier researchers.
Middle AI powers, including the European Union, the Gulf States, India, Japan, and South Korea, hope they can develop an advantage in some part of the AI tech stack—computing power, data centers, energy infrastructure, chips, or talent. This would reduce dependence on Washington and Beijing and help create a multipolar AI world. Those middle powers also hope that diffusing AI capabilities into their economies will matter as much, if not more, than AI innovation. AI sovereignty, which is an attempt by nations to control their own AI stack, has gained traction.
But it remains unclear if the idea is more than a pipe dream for middle powers, much less developing economies. If model development continues to require colossal amounts of energy, computing power, and capital, and if advanced AI models can be applied to AI development, thereby creating systems that can rapidly self-improve, then the gap between the AI powers and the followers could become even wider. In that scenario, middle AI powers will never be able to catch up.
A defining characteristic of the AI age is the overwhelming influence of the private sector. Over the last half century, public funding and research drove discoveries in the technologies that have been critical to U.S. power, including nuclear physics, space exploration, and the internet. Today, the private sector firms Anthropic, Google, and OpenAI are the largest engines of AI innovation. Whereas Washington invested roughly $3.3 billion in nondefense AI research and development in the 2025 fiscal year, U.S. private sector investment in AI exceeded $109 billion in 2024. As researchers from OpenAI observed in February 2026, “It is plausible that private-sector spending aimed at developing and scaling advanced AI capabilities now exceeds the military research and development budgets of nearly every country in the world.”
Tech companies’ power, influence, and reach is growing and becoming comparable, in some respects, to that wielded by countries. As the political analyst Ian Bremmer wrote in Foreign Affairs, “These companies exercise a form of sovereignty over a rapidly expanding realm that extends beyond the reach of regulators: digital space.” Those firms could set technical and safety standards that preempt or constrain government attempts to regulate them at the domestic and international level. They could also cut weaker states off from access to data, compute, or models if there is a dispute over taxation, content moderation, or other regulatory issues.
The leading labs understand the technology that they are building better than the regulators. The result is that governments may be unprepared by dramatic breakthroughs in model capacities or unaware of systemic vulnerabilities, both of which could lead to economic and political dislocation. If the frontier labs do achieve superintelligence, their capabilities will become extraordinary. They could launch cyberattacks, wield political influence, or fashion new and exotic weapons to seize power.
U.S. policymakers need to ensure that there is a productive and predictable balance of power between AI firms and the state. No such balance currently exists, as was shown in the late 2025–early 2026 fallout between Anthropic and the Pentagon over the deployment of AI for surveillance and autonomous weapons. In an apparent bid to punish the company and intimidate other American technology firms, the Pentagon designated Anthropic a supply chain risk, even as its Claude model was being used to support U.S. military operations in Iran. The real question of who should govern AI—the firms or the state—remains to be answered.
Over the last decade, China has developed a comprehensive plan of military-civil fusion, speeding the flow of dual-use technologies like AI to the People’s Liberation Army. Although the interests of the technology firms and the state do not always overlap, there is little question that the Chinese Communist Party is in command, setting long-term strategy and channeling private sector energy toward government goals. The stability of the relationship is one of China’s competitive advantages in the global AI competition.
In response, U.S. policymakers should aim to build strong relations with the firms and maintain a regulatory environment that supports innovation. Do that, and the United States will likely continue to lead the world in AI development. Overstepping or politicizing the relationship, on the other hand, could not only slow innovation, but also cause international partners to mistrust the U.S. tech stack. The effect of that would be to slow the uptake of American AI in the rest of the world.