In a Beijing conference room recently, I noted to a counterpart that the United States and the Soviet Union took more than a decade to build toward nuclear coordination after the Cuban Missile Crisis. "That did not end well for the Soviet Union," was the reply.
As Donald Trump meets Xi Jinping, the headline items are trade, Taiwan, and Iran. Artificial intelligence (AI) is on the agenda, too. Treasury Secretary Scott Bessent named two key risks of powerful AI that demand the leaders’ attention: weaponization and runaway AI.
I met with Chinese Communist Party advisers, AI lab directors, and national security strategists about these very risks. I learned that President Trump's approach to AI and China is working. Every Chinese counterpart I met named the U.S. policy stack. Pax Silica. The Genesis Mission. The White House directive on adversarial distillation. They named legislation by sponsor. The complaints arrived in the same order – export controls, remote compute access restrictions, entity listings – before pivoting to a new request: shift the conversation from competition to managed risk.
Anthropic's limited release of Mythos in April made the stakes real. A model that can find unpatched vulnerabilities across every major operating system is not a commercial product. It is a national security capability. The next model will be more capable, the one after that more capable still.
I went to China hesitantly, as a skeptic of broad cooperation with China. But cooperation on AI is not the goal. The goal is establishing a competition floor. Above the floor, America accelerates as hard as it can. Pax Silica stays untouched. Export controls tighten. We compete without restraint on market share, models, chips, and standards. We play to win. Below the floor, we coordinate on the narrow category of risk that does not respect borders.
Yes, Chinese companies cheat, distill U.S. models, and steal AI chips. But a competitive floor recognizes a shared threat that neither government can solve alone. An AI threat emanating from a Chinese company increases risk for Beijing just as an AI threat from an American company does for Washington. Many officials I met understand this, even as they complained about export controls. Mutual threat is the foundation that holds strong even when there is no trust.
The smart way forward: the presidents leave the summit with a single sentence in a joint statement acknowledging shared AI risk, plus a designated counterpart on each side empowered to build the coordination mechanisms below the competitive floor.
Two years ago today, the Biden administration sat down with Chinese officials in Geneva for the first U.S.-China AI dialogue. It produced nothing because there was no leverage. There is now. The rapid pace of AI development means that this effort has come none too early.
In his opening remarks today, Mr. Xi told Mr. Trump that "the world has come to a new crossroads" and asked whether the two countries can "transcend the so-called Thucydides Trap and pioneer a new paradigm of major-country relations." He called on the United States and China to be "partners, not rivals." That is a false dilemma. Either the two countries transcend the Trap by easing pressure, the framing implies, or they fall into it. There is a third option: a competition floor with guardrails on the narrow set of risks.
Weaponization and runaway AI are exactly that category. Mythos showed the first risk in operational form. The second is no longer theoretical either. American frontier labs now publish safety evaluations in which their own models attempt to deceive evaluators, disable oversight mechanisms, or take unsanctioned actions to preserve their goals. These behaviors appear in controlled tests today. They are already starting to appear in deployed agentic systems.
Answering Mr. Xi with a competitive floor is the strongest position the United States can take. It accepts the seriousness of the risk he raised. It refuses the broader concession he is hoping for, and it rejects the false choice he is offering.
Mr. Trump arrives in Beijing with leverage on the inputs China needs to reach the frontier, and he should use it. Compete without restraint above the floor. Coordinate narrowly below it. Leave with a mandate to develop a framework on AI weaponization and runaway risk that holds regardless of how the rest of the relationship moves. If anyone can achieve this, it is President Trump.
Mark Beall is President of the AI Policy Network. He served as the inaugural Director of AI Strategy and Policy at the Pentagon.