The two leading AI powers are no longer pursuing the same goal. They have, quietly, begun pursuing different goals.
The United States has settled on the phrase super intelligence. The phrase appears in federal communications. It appeared, recently, in citations for the National Medal of Science. It describes a target: a machine intelligence that exceeds human capability across every relevant dimension, capable of recursive self-improvement, capable of acting as a general-purpose autonomous agent in the physical and digital world.
China has begun talking about fusion intelligence. The phrase is less common outside specialist literature, but it has been used by senior policy voices and several leading labs. It describes a different target. Not a machine that surpasses humans. A system in which machines, humans, and the surrounding environment are coordinated so that the resulting intelligence is greater than any of its parts.
The two paths diverge long before any technical decision. They diverge on what intelligence is.
The Super Intelligence Road
The US approach is the more familiar one. Compute and algorithms are the inputs. Scaling them is the strategy. Recursive self-improvement and general autonomous agency are the milestones. The implicit model of intelligence is that it can, in principle, be instantiated entirely inside a machine, and that the right benchmark is human capability in the abstract.
The risks of this approach are well understood inside the field. Sandbox escape. Hallucination. Autonomous action beyond intended scope. The capital intensity of the approach also creates a class dynamic that the strategy does not directly address. A small elite controls the intelligence. A much larger group is replaced by it. The economic and political concentration that follows is a feature of the model's success, not a side effect.
The philosophical premise is also visible. Intelligence is treated as a property that can be owned, scaled, and shipped. It is the kind of intelligence that can be exported, sold, or withheld. It is, in a sense, a commodity.
The Fusion Intelligence Road
The fusion-intelligence framing starts from a different premise. Intelligence is not a property of an isolated agent. It emerges from the interaction between a computational system, the humans who deploy it, and the environment in which both operate. The goal is not to maximise the intelligence of the machine. It is to maximise the intelligence of the whole system.
The strategy has two technical streams. The first, technology-driven AI, continues the work of scaling models, embodied robotics, and perception. The second, demand-driven AI, takes industrial needs, social priorities, public-welfare constraints, and regulatory requirements as inputs to the research and deployment process. Constraints are not obstacles. Constraints are the design specification.
The approach does not give up on frontier capability. It argues that frontier capability, deployed without an understanding of how it will interact with the human and environmental systems around it, is brittle. A model that can solve any problem in principle but causes social disruption in practice is not, in this framing, intelligent. It is just powerful.
The Philosophical Distance
The two framings map onto different philosophical traditions in ways that are not accidental.
Super intelligence is closer to a Western instrumentalist view of technology. The tool is meant to amplify the operator. The benchmark is what the operator can do with the tool. The implied endpoint is a tool so capable that it becomes an operator in its own right.
Fusion intelligence draws from an Eastern philosophical lineage. Confucian attention to human relations and ethical constraints. Daoist sensitivity to environmental conditions and natural limits. Buddhist wariness of treating any single intelligence, human or otherwise, as a fixed point of reference. The implied endpoint is not a single superintelligent agent. It is a sustainable ecosystem in which many different kinds of intelligence contribute to a shared outcome.
Neither framing is a marketing slogan. Both have measurable implications for how research dollars are spent, what kinds of models are deployed, what kinds of products get shipped, and what kinds of failures get tolerated.
The Failure Modes of Super Intelligence
Super intelligence faces a particular set of structural problems. Alignment at human-superhuman capability is harder than alignment at sub-human capability, because the system is increasingly capable of identifying and circumventing the constraints placed on it. Capital concentration concentrates the resulting power in a small number of actors. Displacement of human labour happens faster than social systems can absorb. The strategy's success, taken to its logical conclusion, creates the conditions for its own social backlash.
None of these are reasons the strategy cannot work. They are reasons the strategy requires unusually good governance to deploy safely. The governance infrastructure is, by the strategy's own design, downstream of the technology it is meant to govern.
The Failure Modes of Fusion Intelligence
Fusion intelligence faces a different set of problems. The strategy requires sophisticated coordination across government, industry, and academia. It requires an institutional capacity to translate social priorities into technical constraints. It requires that the people doing the translation understand both the technology and the social context. None of these are easy.
There is also a competitive risk. A strategy that prioritises social fit and ecosystem sustainability may, in the short term, produce systems that are less capable on raw benchmarks than systems that prioritise capability without constraint. The risk is that the more constrained approach loses the race to the more permissive one, and ends up with the worst of both worlds — less capability, less social fit.
What the Divergence Is Not
The split is not a clean ideological one. There are US researchers pursuing fusion-intelligence-style integration problems. There are Chinese researchers pursuing super-intelligence-style scaling problems. The two camps cross-pollinate at every major conference.
The split is also not a race. The leading American labs have not set out to win against China. They have set out to build super intelligence. The leading Chinese labs have not set out to win against the United States. They have set out to integrate AI into the largest manufacturing base and the largest consumer market on earth. The strategies are different because the goals are different.
The Real Test
The test of either strategy is whether the technology, ten years after deployment at scale, has made life better for a meaningful fraction of the population that lives with it. Super intelligence will be tested on whether it produces capabilities that justify the social disruption it creates. Fusion intelligence will be tested on whether its coordinated systems produce outcomes that justify the coordination overhead.
The strategies are different because the underlying bets about intelligence are different. One bet is that machine intelligence can be scaled to a level that makes the rest of the system irrelevant. The other bet is that intelligence has always been a system property, and that the right place to invest engineering effort is the system rather than the component.
Either bet could turn out to be right. The next decade will tell.
What is no longer in doubt is that the two leading AI powers are pursuing those bets on different tracks. The interesting question is no longer which model is bigger. It is which intelligence the world is building.