China has made its view of artificial intelligence unmistakably clear.
Beijing is reportedly preparing to restrict overseas access to its most advanced AI models even as second-tier Chinese firms race to capture global customers that American export controls may drive away. China treats frontier AI as a national asset to guard, not a product to sell freely on the open market.
America’s better bet is disciplined openness: protect frontier capabilities from adversaries while making it easy for allies and trusted partners to build with American AI.
No surprise. Authoritarian governments always seek to control information, technology, and who may benefit from either. The United States should not make the same mistake.
America’s technological advantage has never come from secrecy alone. It comes from building products others want to use, setting standards others trust, and scaling companies faster than any rival.
As Congress negotiates the National Defense Authorization Act, lobbyists are pushing for new AI export restrictions. The impulse is understandable. Advanced chips, large-scale cloud computing, and frontier model weights can help foreign militaries, cyber actors, weapons proliferators, and surveillance states.
Congress must distinguish between two goals: denying adversaries dangerous capabilities and keeping the world’s trusted AI ecosystem centered on the United States.
That distinction should guide every AI export policy.
The first principle is national security through technological leadership. American AI companies, cloud providers, chip designers, universities, and capital markets form the engine of that leadership. Export controls should protect the system, not cripple it.
Second, allies and trusted partners should receive privileged access. America needs an AI coalition, not an AI fortress.
NATO allies and partners such as Japan, South Korea, Australia, Taiwan, and New Zealand should not face the same practical restrictions as China, Russia, Iran, or North Korea when they meet strong security and anti-diversion standards.
Friendly countries will not build around American technology if they believe Washington can abruptly cut off access through broad and discretionary rules. They will build alternatives.
Third, controls should be risk-based and technically precise.
RELATED: The question of the AI age: How much humanism is too much?
A transfer of frontier model weights is not the same as API access to a safeguarded model. A massive training run is not the same as ordinary inference. A Canadian researcher is not a Chinese military-linked laboratory.
The more a transaction transfers durable frontier capability, the stronger the case for licensing. The more it resembles ordinary use of a monitored service, the stronger the case for presumptive access.
The United States already uses focused tools to keep dangerous technologies away from dangerous actors. Those tools work best when they target specific end users and end uses: foreign militaries, sanctioned entities, weapons programs, offensive cyber operations, surveillance abuses, and diversion through shell companies or intermediaries.
They work poorly when companies cannot determine what the law allows.
That is the problem with the Remote Access Security Act, which the Senate is considering.
The remote-access loophole is real. Export controls fail when restricted firms cannot buy advanced chips but can rent the same computing capability through the cloud.
The House-passed bill, however, does more than close that loophole. It allows the Commerce Department to regulate remote access by a foreign person through a network or cloud service whenever the secretary determines that the use could pose a serious risk to U.S. national security or foreign policy.
That grants regulators enormous latitude.
“Foreign persons” include individuals and businesses in friendly countries. “Foreign policy” is an elastic standard. Together, those terms could give the executive branch sweeping authority over global access to American AI infrastructure.
The bill requires Commerce to inform Congress about anticipated regulations, including the security risk, method of regulation, and likely economic consequences. It does not require congressional approval before the rules take effect.
That may be administratively convenient. It is not sufficient for a power that could reshape the global AI market.
Regulators need discretion in a fast-moving technology sector. They do not need unbounded discretion.
The bill’s vagueness would change business behavior long before any court tested its limits.
RELATED: Smartphone and laptop prices are rising. Here’s why they won’t stop.
When companies do not know who may receive access, they overblock. Cloud providers may exclude broad categories of foreign customers. AI laboratories may avoid international partnerships. Universities may slow research collaborations. Trusted allies and legitimate businesses may find themselves treated like hostile actors.
Customers who cannot obtain American AI will find it elsewhere. Increasingly, elsewhere means China.
DeepSeek, Alibaba, Huawei, Tencent, and Kuaishou’s Kling AI are aggressively courting users, developers, and investors worldwide. Every needless barrier Congress places before legitimate users of American AI gives those companies an opening.
The economic stakes are enormous.
PWC estimates that AI could add as much as $15.7 trillion to the global economy by 2030, with China among the largest projected beneficiaries. The Stanford AI Index reports that American private AI investment exceeded $285 billion in 2025, more than 23 times China’s total.
That lead is extraordinary. It is not permanent.
Customers gravitate toward ecosystems that are capable, accessible, reliable, and politically predictable. American dominance will erode if Washington makes access uncertain while Beijing makes its second-tier products easy to obtain.
Congress should not discard the Remote Access Security Act. It should improve it. A better bill would close the cloud loophole while creating safe pathways for lawful use. It should provide privileged access for allies and vetted users, define clear thresholds for frontier-scale compute and controlled model weights, and distinguish among weight transfers, training compute, inference, and ordinary applications.
It should also establish licensing deadlines, appeals or advisory opinions, safe harbors for companies that perform good-faith know-your-customer and anti-diversion checks, and periodic reviews that update or retire controls as technology changes.
That approach is not soft on China. It is how America can confront China without sabotaging itself.
The United States should deny adversaries the tools they need for military modernization, cyber operations, weapons programs, and authoritarian surveillance. It should also make safe transactions easy and dangerous transactions hard.
China is betting that control will win the AI race. America’s better bet is disciplined openness: protect frontier capabilities from adversaries while making it easy for allies and trusted partners to build with American AI.
Congress should write export rules that preserve that advantage, not hand American customers to Beijing by accident.
That model has kept American companies at the front of global technology for decades. It is how they will stay there.
