The Limits of US–China AI Cooperation
A question that refuses to sit still
Every few weeks, the same question returns in a slightly different form. The United States and China hold talks on artificial intelligence. Then the talks produce little that looks like a breakthrough. Washington tightens restrictions on advanced technology. Beijing accelerates its drive for technological self reliance. Chinese open weight models spread through the global developer ecosystem. American officials warn about national security risks. Then, almost in the same breath, representatives from the two countries begin discussing AI safety, crisis communication or the possibility of common guardrails.
Follow only the headlines and the relationship can appear almost self contradictory: cooperation and confrontation, integration and decoupling, dialogue and technological containment, sometimes all in the same month. That is why the apparently simple question, will the United States and China eventually combine their efforts on AI?, needs a more careful answer.
My short answer is this: the United States and China are very unlikely to build a genuinely shared AI ecosystem during the second half of the 2020s. But limited cooperation on safety, crisis management and a small number of governance questions is not only possible; parts of it are already emerging. Strategic alignment between Washington and Beijing on AI, however, remains a very different proposition.
The key is to stop treating “cooperation” as a single thing. The original question contains several different questions, and their answers are not the same.
What would “cooperation” actually mean?
When politicians, journalists and technology executives talk about US China AI cooperation, they often use the same word to describe fundamentally different relationships. There are at least four.
The first and deepest form would be joint research and development. American and Chinese frontier laboratories would collaborate on model architectures, share datasets or training techniques, coauthor important papers, develop common benchmarks and perhaps even operate shared research programmes. This would be genuine scientific integration, and it is also the least plausible form of cooperation in the present strategic environment.
The second is commercial interdependence. American semiconductor, software and cloud companies sell products into China; Chinese developers build systems incorporating American technologies; investment and licensing flow across borders; and firms on both sides have financial reasons to keep parts of the relationship open. This already exists, but under increasingly elaborate political and regulatory constraints.
The third is scientific and Track II dialogue. Researchers, former officials, academics and safety specialists meet outside formal diplomatic negotiations to discuss model evaluations, military risks, loss of control scenarios, cybersecurity, autonomous systems and potential rules of the road. This kind of contact is comparatively resilient because the participants do not need to solve the entire geopolitical relationship before speaking to one another.
The fourth is formal intergovernmental governance: negotiated agreements, crisis communication mechanisms, shared understandings about military AI, incident notification procedures, safety testing principles or, in the most ambitious version, an international system for supervising advanced AI.
These four categories matter because evidence of progress in one says remarkably little about the others. Washington and Beijing can simultaneously restrict technology transfers, accuse one another of irresponsible behaviour and still conclude that it would be useful to have a telephone number to call if an AI system triggers a serious military or infrastructure incident. That is not hypocrisy. It is how strategic rivals behave when the cost of misunderstanding becomes sufficiently high.
Why deep AI integration remains so difficult
The central obstacle is not a lack of scientific curiosity. It is that artificial intelligence has become entangled with national power. Advanced AI is no longer treated merely as another commercial technology. It touches military planning, intelligence, cybersecurity, industrial policy, scientific research, surveillance, communications and economic productivity. The most capable models increasingly depend on large concentrations of computing power, specialised semiconductors, sophisticated networking equipment and enormous capital expenditure.
That gives governments strong reasons to care about who controls every layer of the stack, and the United States and China already compete across many of those layers.
The semiconductor divide
Semiconductors remain the clearest example. Beginning in 2022, the United States progressively tightened restrictions on China’s access to some advanced computing chips and semiconductor manufacturing technologies. The Biden administration then introduced its broader AI Diffusion Rule in January 2025.
The Trump administration rescinded that rule in May 2025, arguing that it was too restrictive toward friendly countries, but the rescission did not mean that Washington abandoned controls aimed at China. The Commerce Department simultaneously emphasised restrictions intended to prevent advanced American computing technology from strengthening Chinese military and AI capabilities.
Since then, the policy has become more transactional rather than simply disappearing. Nvidia disclosed that licences granted beginning in February 2026 permitted limited shipments of H200 processors to specified Chinese customers, subject to inspection in the United States and a 25 percent tariff generated by the requirement that the chips pass through US territory. Nvidia also reported that Chinese restrictions limited its ability to complete the authorised sales.
That is a revealing arrangement because it is neither full technological separation nor normal commerce. It is controlled interdependence.
The cloud problem
There is also a broader question: if advanced chips cannot legally be exported to certain users, should those users be able to access the same computing capability remotely through overseas data centres?
A bipartisan proposal introduced by Senators Dave McCormick and Ron Wyden in December 2025, the Remote Access Security Act, sought to extend US export control authority to certain forms of remote access to controlled technology. Its logic is straightforward. Export controls lose some of their meaning if controlled computing power can simply be rented across a border rather than physically shipped across it.
Whether every proposal becomes law is less important than what the debate reveals: compute itself is increasingly treated as a strategic resource. That makes shared frontier infrastructure between the United States and China exceptionally difficult to imagine.
Two competing technological ecosystems
China, meanwhile, has continued building a domestic regulatory and technological environment around AI. Its rules already include algorithm filing and security requirements, and new regulations governing the labelling of AI generated content took effect in September 2025. Those rules require various forms of explicit or embedded identification for synthetic material and sit alongside China’s broader framework for generative AI services.
The American system remains much more fragmented and market oriented. The contrast does not mean that one country regulates AI while the other does not. Both regulate it. But they regulate different things, through different institutions, toward different political objectives.
That makes interoperability difficult. The result is not a clean technological Iron Curtain. It is something messier: two increasingly distinct strategic ecosystems connected by selectively controlled commercial, scientific and open source channels.
Then came DeepSeek
Any account of the US China AI relationship after 2025 has to reckon with DeepSeek. DeepSeek R1 was released in January 2025 under an MIT licence, allowing unusually broad reuse and modification. DeepSeek presented the model as achieving performance comparable to OpenAI’s o1 on several reasoning tasks.
The significance went beyond one model. DeepSeek challenged a comforting assumption in parts of Washington: that restricting access to the most advanced American chips would automatically translate into a large and durable American advantage in frontier AI.
Export controls may still raise costs and constrain scale. But DeepSeek demonstrated that Chinese laboratories could remain formidable competitors by improving efficiency, exploiting available hardware and disseminating models aggressively. It also strengthened a second Chinese advantage: open weight distribution.
Instead of requiring every user to access a model through a centrally controlled API, an open weight model can be downloaded, modified, fine tuned and deployed on infrastructure controlled by the user. That changes the geography of technological power.
Once a capable model’s weights are released, its influence is no longer confined to the country that produced it. Developers in Jakarta, São Paulo, Berlin, Riyadh or California can build on it. China therefore does not need to persuade every country to buy a vertically integrated Chinese AI stack. It can gain influence simply by making useful models easy and inexpensive to adopt.
By September 2026, Chinese open weight models had become major components of the global AI ecosystem. Various analyses showed Chinese developed open models accounting for a substantial share of usage on platforms such as OpenRouter, even while proprietary American systems continued to capture much of the commercial revenue.
That is an important distinction. Adoption is not the same as revenue, and open weight influence is not the same as commercial dominance. But influence still matters.
The United States rediscovered open weights
For several years, the most prominent American frontier AI companies moved increasingly toward closed systems. Then the strategic environment changed.
OpenAI released GPT OSS 120B and GPT OSS 20B in August 2025, its first open weight language models since GPT 2. The models were released under the permissive Apache 2.0 licence and designed to run outside OpenAI’s own infrastructure.
The decision reflected a broader debate inside the United States. Closed models allow companies to preserve intellectual property, maintain control over access, update safeguards and monetise usage. Open weight models offer something different: diffusion. They allow researchers, companies and governments to adapt models themselves. They encourage ecosystems of applications and derivatives. Once developers have built infrastructure around a model family, that ecosystem can become strategically important.
So the US China AI contest is no longer simply a contest over who possesses the single most powerful model. It is also a contest over which technological ecosystem other people choose to inhabit.
The American model has historically benefited from extraordinary private capital, leading semiconductor companies, global cloud providers and frontier laboratories. The Chinese model increasingly combines domestic industrial policy with inexpensive, widely distributed model weights. Neither strategy has conclusively defeated the other, and that makes global AI development much harder to divide neatly into American and Chinese blocs.
The paradox: geopolitical rivalry, technological entanglement
This is where the simple “decoupling” story begins to fail. Governments can restrict semiconductor exports, screen investments, prevent sensitive laboratories from collaborating and restrict government procurement. What they cannot easily do is prevent ideas from crossing borders.
Machine learning papers are read internationally. Researchers follow each other’s work. Open model weights circulate globally. Evaluation techniques are reproduced. Software libraries are copied. Developers test competing models through the same platforms. An architecture proposed in Beijing can influence a team in San Francisco within days, just as an American evaluation benchmark can be used by a laboratory in Hangzhou.
A Chinese open weight model can be deployed by an American startup without the two organisations ever signing a cooperation agreement. This is not diplomatic cooperation, but neither is it technological separation.
AI is unusually difficult to contain because much of its value resides not in a single physical object but in knowledge: algorithms, weights, training recipes, research insights, software and engineering practices. Semiconductors can be stopped at customs. Ideas do not pass through customs in quite the same way.
That makes the emerging US China AI relationship fundamentally different from the Cold War competition between the United States and Soviet Union. The Soviet Union never had millions of Western programmers downloading Soviet software from public repositories. AI competition is occurring inside a digitally interconnected world, so the rivalry is likely to remain porous by design and leaky in practice.
Where cooperation is already real
If deep integration is unlikely, it does not follow that cooperation is absent. It exists primarily where both sides can identify a danger that neither can comfortably manage alone.
Government to government dialogue
The first formal US China intergovernmental talks devoted specifically to artificial intelligence took place in Geneva in May 2024. The US State Department said beforehand that the delegation would discuss risks arising from advanced forms of AI and described maintaining communication as part of “competing responsibly.”
That phrase captures much of what has followed. The objective was not partnership in developing AI. It was managing the consequences of competition. The process remained fragile, but it established a precedent: Washington and Beijing could treat AI risk as a subject worthy of direct bilateral discussion even while disagreeing profoundly about technology policy.
Human control over nuclear decisions
Another significant moment came from the broader Biden Xi dialogue in 2024, when the two governments publicly affirmed the importance of maintaining human control over decisions involving nuclear weapons. The commitment was narrow. It did not constitute an AI arms control treaty, nor did it create an inspection regime.
But narrow agreements are exactly what one should expect at the beginning of a relationship like this. The important principle is that neither government wants ambiguity about whether an autonomous system could initiate a catastrophic escalation.
Track II engagement
Unofficial dialogue has often gone further than formal diplomacy. Researchers and former officials from the United States and China have maintained Track II conversations on AI and national security through periods of broader tension. Those exchanges have often been sustained by mutual interest rather than mutual trust, particularly around the possibility that AI could increase the risk of accidental escalation.
That distinction is essential because cooperation between geopolitical rivals rarely begins with trust. Sometimes it begins because both sides distrust each other so much that they need mechanisms for preventing misunderstanding.
Multilateral governance
China has also continued participating in multilateral AI governance processes. At the Paris AI Action Summit in February 2025, more than sixty countries, including China, supported a declaration calling for AI that was open, inclusive, transparent, ethical, safe, secure and trustworthy. The United States and United Kingdom declined to sign it.
That does not mean China and the signatories suddenly shared one regulatory philosophy. International declarations are often deliberately broad enough to accommodate competing political systems. But participation itself matters because Beijing wants influence over the rules governing global AI, and Washington wants influence over those rules too.
Global governance is therefore becoming another arena of competition, but competition inside an institution is different from refusing to participate in the institution at all.
The Beijing summit of May 2026
The most important recent development occurred during President Donald Trump’s meeting with President Xi Jinping in Beijing in May 2026. Following the summit, the two governments agreed to establish a governmental dialogue on artificial intelligence. Subsequent analysis described the emerging channel as focused on relatively narrow questions of AI risk and governance rather than resolving the strategic technology competition itself.
By September 2026, US and Chinese officials were also discussing the possibility of mechanisms through which the two countries could communicate about major AI incidents with national security consequences. No agreement should be assumed before one exists, but the direction is noteworthy.
The emerging model of cooperation is therefore not “let us build AI together.” It is closer to: “If something goes badly wrong, how do we prevent each other from misunderstanding what happened?” That is a much narrower ambition, but it may also be a much more realistic one.
AI cooperation may end up looking more like arms control than CERN
There are two very different models for international scientific cooperation. One is CERN, where governments pool money, infrastructure and scientific talent because their objectives are broadly aligned. Scientists collaborate openly on a shared project.
That is not a useful model for US China frontier AI relations.
The second model is Cold War arms control. The United States and Soviet Union remained strategic rivals. They did not merge their nuclear laboratories or share their most sensitive weapons designs. Yet over time they created hotlines, notification arrangements, arms control agreements and norms designed to reduce the probability of accidental catastrophe.
The reason was not friendship. It was mutual vulnerability, and AI may develop in a similar direction.
Imagine an advanced autonomous cyber system behaving unpredictably inside another country’s critical infrastructure. Was it an attack? Was it an accident? Was it deployed by a state, or created by a criminal group? Has the government that developed the underlying model lost control of it?
Now imagine that the incident affects an electrical grid, communications network, financial market or military early warning system. The technical problem immediately becomes geopolitical.
A hotline matters because attribution takes time while political leaders may feel pressure to respond immediately. An incident reporting mechanism matters because silence can be interpreted as hostility. Shared terminology matters because two governments may otherwise use the same words to describe very different thresholds of danger.
None of this requires the United States and China to agree about who should dominate the AI industry. It requires only that both prefer controlled competition to uncontrolled escalation. That is a much lower threshold for cooperation.
Why even narrow cooperation remains difficult
The case for cooperation should not be confused with an assumption that it will be easy. Several structural obstacles remain.
Different threat priorities
American debates about frontier AI often devote substantial attention to biological misuse, cyber capabilities, autonomous systems and the possibility that increasingly capable AI could become difficult to control. Chinese governance has historically placed greater institutional emphasis on information integrity, social stability, political security, content control and state supervision, while Chinese researchers and officials have increasingly also discussed frontier and loss of control risks.
These categories overlap, but they are not identical. Two governments can agree that AI is dangerous while disagreeing over what the danger actually is. That makes negotiation difficult because an agreement requires more than shared anxiety. It requires compatible definitions.
Different governance systems
The United States regulates AI through a mixture of federal agencies, state governments, sector specific rules, procurement requirements and private standards. China uses a more centralised regulatory framework involving algorithm registration, security obligations, content rules and government oversight.
Those systems reflect broader differences in political institutions. A comprehensive mutual recognition regime would therefore be extraordinarily difficult. Even if an American model satisfies American requirements, Beijing may have separate political, security and content requirements. Even if a Chinese model satisfies Chinese regulation, Washington may still see national security risks in deploying it inside sensitive American systems.
Verification
Verification may be the hardest problem of all. Traditional arms control agreements can sometimes count physical objects: missiles, launchers, warheads or test sites. AI capability is less tangible.
How do inspectors determine whether a laboratory has trained a model above an agreed capability threshold? How do governments verify that a supposedly civilian model has not been adapted for military use? What counts as a frontier model when capabilities improve continuously? How should fine tuning be treated? What about distributed training? What about a model whose weights are already public, or a system whose dangerous capabilities emerge only after deployment?
These questions do not make governance impossible. They do explain why highly ambitious AI treaties are much harder than political speeches sometimes suggest.
Why both countries still have reasons to cooperate
Despite those obstacles, there are powerful reasons for Washington and Beijing to maintain at least a narrow safety relationship.
AI incidents will not respect borders
A serious AI enabled cyberattack does not become harmless when it crosses the Pacific. A dangerous biological capability produced by a model does not remain a national problem, and an autonomous system that behaves unexpectedly inside a military crisis can affect both sides.
A widely released open weight model can also spread globally before any government can recall it. The more powerful AI becomes, the less realistic it is to assume that one country’s safety failures will remain geographically contained.
Open weight models create irreversible diffusion
This is particularly important. A closed AI service can theoretically be updated, monitored or switched off by its provider. An open weight model cannot easily be recalled once millions of copies exist.
That produces a shared problem. If highly capable models become widely downloadable, every leading AI power acquires an interest in understanding how other countries evaluate dangerous capabilities before releasing them.
That does not necessarily imply a treaty. It could begin with something much smaller: shared evaluation terminology, voluntary incident disclosure, red team exchanges or communication between national AI safety institutions.
Critical infrastructure is mutually vulnerable
The United States and China are both highly digitised major economies. Banking, electricity, logistics, communications, manufacturing and transportation increasingly depend on software. AI increases both the defensive capabilities available to these systems and the potential sophistication of attacks against them.
Neither government benefits from a world in which every unexplained AI driven infrastructure failure risks being interpreted as state aggression. Crisis communication is therefore not altruism. It is insurance.
Neither government wants an AI system it cannot control
Strategic rivalry can obscure one surprisingly basic common interest: Washington wants powerful AI systems to remain controllable, and Beijing wants powerful AI systems to remain controllable.
The political reasons may not always be identical. The institutions used to achieve control may differ. The acceptable level of risk may differ. But an advanced system behaving autonomously in ways that neither government can predict is not an attractive scenario for either capital.
That narrow shared interest may ultimately prove more important than grand declarations about technological partnership.
The most plausible future: competitive coexistence
So what should we actually expect between now and 2030? Not technological union, but not complete separation either. The most plausible outcome is something I would call competitive coexistence.
The United States and China will continue racing to improve models, secure semiconductor supply chains, build data centres, attract scientific talent and persuade other countries to adopt their technologies. Washington will continue treating some advanced technologies as national security assets. Beijing will continue investing heavily in technological self sufficiency. Chinese developers will continue competing for global adoption, while American companies defend their position at the frontier.
And yet, running alongside that competition will probably be a thinner network of communication mechanisms. There may be regular governmental AI dialogues, agreements about human control over particularly sensitive military systems, voluntary notification arrangements for major incidents and discussions about cybersecurity, biological misuse and critical infrastructure.
Researchers will continue talking across borders even when official diplomacy deteriorates. Open weight models will continue making technological separation incomplete. Commercial transactions will periodically reopen where governments decide the economic benefits exceed the security costs.
This will look inconsistent if we insist on putting the relationship into one box. It becomes much easier to understand once we stop trying.
My working forecast for 2026 to 2030
I would divide the possibilities into several broad bands rather than pretending we can predict diplomatic outcomes with precision.
Continued bilateral and Track II AI dialogue is highly plausible. The institutional foundations now exist. The Geneva talks established formal contact, Track II channels have persisted, and the May 2026 Beijing summit revived governmental engagement. Unless the broader political relationship deteriorates dramatically, continued dialogue is easier to imagine than its complete disappearance.
A bilateral mechanism for reporting serious AI incidents is plausible. This has moved from an abstract policy idea toward a subject of bilateral discussion. Whether any current proposal becomes a formal mechanism remains uncertain, but the logic is strong because a communication mechanism does not require either country to surrender meaningful technological advantage.
Limited commercial relaxation in exchange for concrete concessions is also plausible, but unstable. The history of licensing decisions around advanced chips demonstrates that policy can move in both directions depending on the broader trade relationship. It is therefore better understood as an instrument of statecraft than as a linear march toward either total embargo or complete liberalisation.
A binding bilateral AI arms control treaty is considerably harder. The verification problems are formidable, the technology evolves rapidly, and the two governments do not yet appear to share enough assumptions about governance to support an extensive treaty architecture. Smaller agreements are far easier to imagine.
Joint US China frontier laboratories are difficult to envisage under present conditions. This is where national security concerns cut deepest. Sharing advanced training methods, compute, datasets and frontier capabilities would require a level of strategic trust that does not currently exist. Scientific contact will continue, but institutional integration is another matter.
A unified global AI authority jointly led by Washington and Beijing is also difficult to imagine without a major external shock. International institutions may emerge, and both countries will seek influence over them, but a body possessing intrusive authority over frontier development would require governments to surrender information and sovereignty they currently treat as strategically sensitive.
The practical path is therefore more likely to involve overlapping institutions, national regulators and bilateral communication mechanisms than a single world AI regulator.
What could change the trajectory?
Forecasts about AI should be held lightly because the technology itself can change the political environment. Three kinds of shock could alter the relationship rapidly.
A major international AI incident
The most obvious catalyst would be an event demonstrating that existing national safeguards are insufficient. Imagine an AI enabled cyberattack causing prolonged infrastructure failure across several countries, a model materially assisting a biological attack, or an autonomous military system behaving unpredictably during an international crisis.
An event of that magnitude could do for AI governance what earlier strategic crises did for nuclear risk reduction: turn abstract danger into a concrete political problem. Governments frequently accept constraints after disasters that they considered unnecessary before them.
A severe US China military crisis
The opposite shock is equally possible. A confrontation involving Taiwan or another major military flashpoint could rapidly overwhelm the fragile cooperative architecture.
Academic exchanges could be restricted, technology controls could expand, commercial relationships could contract, and AI safety dialogue might itself become politically toxic. The existence of technical cooperation should therefore never be mistaken for insulation from the wider strategic relationship. It depends on that relationship remaining competitive without becoming openly military.
A fundamental change in AI capabilities
The third possibility comes from the technology itself. If models remain powerful but fundamentally manageable tools, governments may continue treating AI primarily as an economic and strategic asset.
If evidence instead emerges that much more advanced systems introduce genuinely global risks that no country can contain independently, national incentives could change. The conceptual model might shift from “AI as the next strategic industry” toward “AI as a transnational risk requiring joint management.”
Climate change offers one analogy. Nuclear weapons offer another. Neither is perfect. The political future of AI will depend partly on which analogy begins to feel more accurate to policymakers.
What American readers should understand about China
There is a temptation in technology competition to underestimate an opponent until it produces something surprising, then overestimate it immediately afterward. Neither response is useful.
China possesses an enormous technology sector, a large engineering workforce, major internet platforms, substantial manufacturing capacity and a government willing to treat artificial intelligence as a strategic priority. At the same time, constraints on access to some advanced semiconductors and manufacturing technologies are real.
The correct conclusion is therefore neither “export controls have stopped Chinese AI” nor “export controls are completely useless.” Both claims are too simple. Controls can increase costs without freezing progress, while Chinese laboratories can innovate around constraints without making those constraints irrelevant.
The competition is dynamic. The lesson of DeepSeek is not that hardware suddenly stopped mattering. It is that hardware is only one variable in an equation that also includes algorithms, engineering, data, capital, energy, talent and distribution.
Open weight distribution also gives Chinese developers a route to global influence even where Chinese cloud infrastructure or hardware is not dominant.
What Chinese readers should understand about the United States
The reverse misunderstanding also occurs. American restrictions on advanced technology are not simply a temporary disagreement that will necessarily disappear when diplomatic relations improve. A large part of Washington now views leadership in advanced computing as directly connected to military and economic power.
That position extends well beyond a single administration. Individual controls may change, licensing arrangements may loosen or tighten, and governments may bargain. But the strategic importance attached to advanced semiconductors, data centres, frontier AI and related technologies is unlikely to disappear.
At the same time, the United States is not a single actor. Its frontier laboratories, cloud companies, semiconductor manufacturers, universities, Congress, executive branch, state governments and national security agencies frequently have different incentives.
A chip company may want access to the Chinese market while a security agency wants tighter restrictions. A university may want scientific exchange. A frontier AI laboratory may simultaneously seek Chinese customers, worry about model extraction and support dialogue on catastrophic risk.
Those tensions are part of American policymaking, not evidence that no policy exists. Understanding that pluralism is essential to understanding why US policy can sometimes appear inconsistent.
The deeper struggle is over the architecture of the AI world
Ultimately, the most important competition may not be over which country produces the highest scoring model in any particular month. It may be over the architecture surrounding AI.
Who manufactures the chips? Who operates the cloud? Whose models do developers build on? Whose technical standards become international standards? Whose safety evaluations are considered legitimate? Which countries host the data centres? Which languages receive first class model support? Which legal frameworks govern deployment? Which payment systems, software libraries and developer platforms sit around the models?
And perhaps most importantly: which ecosystem becomes easier for the rest of the world to join?
The United States has enormous advantages in private capital, frontier laboratories, cloud computing and advanced semiconductor design. China has enormous advantages in manufacturing scale, state coordination, a vast domestic market and an increasingly influential open weight ecosystem.
The rest of the world is not simply watching. Countries in Europe, the Gulf, Southeast Asia, Africa and Latin America will make their own choices, and many will resist being forced into a binary technological bloc.
That means the AI order may become multipolar even if the frontier competition remains primarily American and Chinese.
The central paradox
This brings us back to the apparent contradiction with which we began. Are the United States and China decoupling? Yes, in strategically sensitive areas. Are they still technologically entangled? Yes. Are they competing for global AI leadership? Clearly. Are they beginning to discuss common safety problems? Also yes.
All four statements can be true simultaneously, and increasingly they are.
The mistake is assuming that international relationships must be either cooperative or adversarial. Great power relationships rarely work that way. Countries compete where advantage matters, trade where profit matters, negotiate where instability becomes dangerous and cooperate where the costs of refusing to cooperate become greater than the political costs of talking.
Artificial intelligence does not abolish that logic. It intensifies it.
Closing thought
When the modern AI competition first began to take shape, it was tempting to imagine another space race: two rival systems, two separate technological programmes, a visible finish line and an eventual winner.
That analogy now looks increasingly inadequate. The AI contest is taking place on a global internet, through multinational semiconductor supply chains, across open research communities, inside globally distributed software ecosystems and through models that can sometimes be copied around the world in hours. At the same time, governments increasingly treat the underlying compute, infrastructure and frontier capabilities as matters of national security.
The result is neither a clean Cold War nor a global scientific commons. It is something stranger: the United States and China are constructing rival AI ecosystems while remaining partially embedded in the same technological civilisation. They are trying to reduce dependence on each other while their researchers continue learning from each other. They are restricting access to computing power while open model weights cross borders almost frictionlessly. They are competing to shape the global rules while slowly discovering that some rules may have to be discussed together.
So will the United States and China combine their efforts on artificial intelligence? Not in the sweeping sense the question usually implies. They are unlikely to pool their frontier laboratories, jointly develop their most capable models or abandon the technological competition that now sits at the centre of their strategic relationship.
But they may increasingly cooperate in a narrower and, in some ways, more important domain: preventing the competition itself from becoming catastrophically unstable.
The analogy I keep returning to is therefore not CERN, but the hotline: not shared laboratories, but shared warnings; not a common technological project, but mechanisms for preventing accidents; not trust, but communication despite distrust.
As AI systems become more capable, that modest form of cooperation may turn out to be less dramatic than the optimists hope and more consequential than the pessimists expect. The future of US China AI relations will probably be defined by that uncomfortable combination: competition without complete separation, cooperation without strategic alignment, and interdependence without trust.
It is not a tidy answer. But increasingly, it looks like the right one.
