China Refuses to Slow AI as Trump Calls Jensen Huang Live: Is the Global Safety Debate Becoming an Arms Race?
China rejected calls to contain its AI progress while Trump phoned Nvidia’s Jensen Huang onstage. Safety, industrial policy and geopolitical rivalry are colliding.
China says it will not accept efforts to slow its artificial-intelligence development. Almost simultaneously, President Donald Trump phoned Nvidia chief Jensen Huang during the All-In Summit, dismissed fears of robots taking control and praised data centers. The spectacle captured a widening divide: can governments slow advanced AI for safety when each fears that restraint will hand strategic advantage to the other?
The immediate debate followed an intervention by Anthropic CEO Dario Amodei, who argued for coordinated limits on the pace of frontier-model development and stronger measures to prevent China from acquiring the most advanced chips. Chinese officials and state media called the framing fearmongering and an attempt to contain the country.
Trump’s live call strengthened the opposing message. Huang placed the president on speakerphone before the audience. Trump described AI takeover concerns as a “hoax” and emphasized the economic benefits of infrastructure investment. Huang argued that AI should spread across industries and countries rather than remain concentrated in a few companies.
Neither camp is entirely simple. China is accelerating models, chips, robotics and power infrastructure, but President Xi Jinping has also said AI must remain under human control. Chinese security officials worry that foreign models could spread propaganda or undermine political stability. Beijing supports governance—on terms compatible with state authority and Chinese competitiveness.
American technology leaders likewise disagree. Amodei warns about autonomous systems, biological misuse, cyberattacks and military power. Other executives fear that a pause would be unenforceable, drive development underground or allow less cautious actors to advance. Companies also possess obvious commercial incentives: regulation can protect the public, but complex compliance may entrench firms already rich enough to satisfy it.
The China question makes every safety proposal harder to evaluate. Export controls on advanced Nvidia chips can slow Chinese training capacity, protect U.S. military advantages and reduce misuse. They can also encourage China to develop domestic substitutes, fragment global standards and exclude Chinese researchers from cooperative safety work.
Beijing’s open-source models complicate the market. Western startups can adapt comparatively inexpensive Chinese systems, reducing dependence on major American laboratories. Restricting those models might address security or censorship concerns, but it could also increase costs for smaller companies and strengthen the very U.S. incumbents lobbying for rules.
The data-center debate is equally tangible. Rapid construction creates jobs, tax revenue and computing capacity, yet consumes enormous electricity and water. Communities may welcome investment while opposing power-price increases, noise or land use. Calling every objection anti-technology avoids the need to allocate costs transparently.
Safety advocates are correct that competition does not eliminate catastrophic risks. Nuclear states created arms-control mechanisms precisely because rivals recognized mutual danger. AI systems capable of autonomous cyber operations, weapons assistance or strategic deception could require monitoring, incident reporting and shared red lines.
Acceleration advocates are correct that vague existential warnings can become political tools. A rule written by three U.S. laboratories should not automatically govern the world. Independent evaluation, transparent thresholds and antitrust safeguards are necessary if “safety” is not to become another name for market closure.
Trump’s assurance that strong leadership is sufficient offers political clarity but little technical detail. Liability law can address some harms after they occur; it may not prevent a rapidly replicating cyber capability or loss of control. China’s promise not to slow down similarly leaves unanswered how it will verify that powerful systems remain controllable.
A workable international approach might focus on specific capabilities rather than nationality: mandatory testing for autonomous replication, biological design, cyber exploitation and weapons integration; protected channels for reporting incidents; and verification of the largest training runs. That would still be difficult, but more concrete than asking countries to trust competitors’ intentions.
A further complication is that slowing model training does not necessarily slow deployment. Existing systems can already be adapted for surveillance, misinformation, targeting and cyber operations. Governments may negotiate limits on the next generation while current tools spread through militaries and companies. That argues for a dual approach: frontier evaluations for future models and enforceable rules for present uses. Otherwise, a grand debate about superintelligence may distract from measurable harms—and from the concentration of data, chips and electricity in a small number of firms today.
What to watch next
Will the Trump–Xi agenda include test standards or only chip restrictions? Do Anthropic, OpenAI, xAI, Google and Chinese laboratories agree on measurable danger thresholds? Can smaller developers participate without prohibitive costs? And is the world building an AI safety regime—or disguising a technology arms race inside the language of protecting humanity?