WASHINGTON/BEIJING / The global race for artificial intelligence supremacy is entering a far more consequential phase, with the United States and China increasingly divided not only over who will lead the technology, but over who will write the rules governing systems that could reshape economies, national security, warfare and potentially the balance of global power.

The debate has intensified dramatically in recent days following extraordinary warnings from some of the people building the world’s most advanced AI systems. At the same time, Washington and Beijing are accusing each other of using concerns about safety as cover for a much broader technological struggle involving semiconductors, computing power, cybersecurity, military capabilities and geopolitical influence.

Anthropic CEO Dario Amodei brought the issue to the forefront over the weekend by calling on leading AI companies to deliberately moderate the pace at which they increase the capabilities of their most powerful models. Amodei argued that the industry should buy itself more time to develop safeguards before artificial intelligence advances beyond humanity’s ability to reliably supervise it.

His proposal does not call for stopping AI research. Instead, it would introduce independent safety evaluators inside frontier AI companies, encourage leading developers to coordinate on minimum safety standards and seek broader international cooperation on systems capable of causing catastrophic harm. OpenAI CEO Sam Altman and xAI founder Elon Musk publicly expressed support for significant parts of Amodei’s proposal.

Amodei’s intervention came after a series of incidents and internal warnings raised uncomfortable questions about whether the companies pushing AI forward can always control what increasingly autonomous systems do.

Anthropic researcher Jacob Coxon resigned last week and publicly warned that people closest to frontier AI development genuinely fear future systems could become catastrophically dangerous. Coxon, who previously worked at OpenAI, said competitive pressure could eventually force laboratories to cut corners or weaken oversight as companies and governments race to stay ahead. His departure became particularly significant because Anthropic has long presented itself as one of the industry’s most safety-focused laboratories.

The warnings are no longer limited to theoretical discussions about distant artificial general intelligence. AI systems are already becoming dramatically more capable in cybersecurity, autonomous computer use and complex problem solving.

OpenAI disclosed this month that its GPT-6 Astra model had reached what the company defines as a “Critical” cybersecurity capability level under its Preparedness Framework. OpenAI said Astra can, under appropriate conditions and with access to tools, discover previously unknown vulnerabilities and develop techniques for exploiting hardened systems without continuous human direction. The company delayed portions of Astra’s development while strengthening security controls and ultimately released the model with additional safeguards and restrictions on its most advanced cyber capabilities.

OpenAI had already temporarily slowed some scaling and training activities in August while reinforcing internal containment, monitoring and network-security measures. The company said safeguards had to advance alongside model capabilities rather than fall behind them.

But what might appear to be a technical safety discussion inside Silicon Valley has quickly collided with the geopolitical rivalry between Washington and Beijing.

Amodei has simultaneously argued that the United States should maintain tight restrictions on China’s access to the most sophisticated AI chips and should combat efforts by Chinese laboratories to obtain capabilities from American models through techniques such as model distillation. He has also acknowledged an uncomfortable dilemma: a slowdown by American laboratories becomes strategically difficult if China continues advancing without comparable restraints.

China has responded sharply.

On Monday, Chinese Foreign Ministry spokesperson Guo Jiakun said international cooperation—not confrontation—should guide artificial intelligence development. Beijing warned that promoting threat narratives and engaging in “malicious competition” could disrupt global AI governance rather than make the technology safer.

China’s state-backed Global Times went considerably further, portraying Amodei’s proposals as part of a “Cold War” strategy designed to preserve American technological dominance.

The newspaper argued that restricting China’s access to advanced chips while excluding Chinese companies from international AI governance would create a technological monopoly rather than a genuine safety regime. It said isolating China from frontier AI development could actually increase global risks by fragmenting standards and reducing cooperation between the world’s two leading AI powers.

That criticism highlights one of the central contradictions in today’s AI debate.

Many American technology leaders want slower and more cautious development—but they do not want the United States to slow down so much that China overtakes it.

Chinese officials, meanwhile, support international AI safety mechanisms—but reject rules they believe are designed in Washington primarily to constrain China’s technological rise.

President Donald Trump made the American strategic calculation explicit on September 13.

Asked about calls to slow AI development, Trump dismissed some of the catastrophic scenarios being raised by researchers as exaggerated and described those promoting them as “very negative forces.” He said the United States currently leads China in artificial intelligence and wants to keep that advantage.

“Whoever wins AI wins,” Trump said.

That statement captures why AI safety negotiations could prove extraordinarily difficult.

Artificial intelligence is no longer simply about chatbots answering questions or software helping office workers write documents. Frontier models increasingly intersect with cybersecurity, intelligence analysis, autonomous systems, scientific discovery, military planning, biotechnology, robotics, financial markets and industrial production.

Possession of the most advanced AI could therefore translate into economic, military and geopolitical power.

Washington has spent years restricting China’s access to cutting-edge semiconductors and chip-manufacturing equipment, particularly technologies needed to train enormous AI models. The objective has been to prevent Beijing from easily acquiring the computing infrastructure required to match the most advanced American laboratories.

Yet Chinese companies continue to narrow parts of the gap.

DeepSeek, Alibaba, Moonshot AI, Z.AI and other Chinese developers have increasingly emphasized efficient and open-weight models that can be deployed at lower cost and, in many cases, modified by outside developers. Beijing has promoted that ecosystem as an alternative to expensive and more tightly controlled American platforms.

Another dispute exploded on September 8 when the U.S. government accused six Chinese AI companies—including DeepSeek, Moonshot AI and Alibaba—of using American AI systems to accelerate development of their own models through large-scale distillation.

Distillation itself is a legitimate and widely used machine-learning technique in which knowledge from a larger model helps train a smaller or more efficient system. American officials, however, alleged that Chinese firms were systematically extracting capabilities from models created by OpenAI, Anthropic, Google and other U.S. developers.

China rejected the accusations, describing them as unsupported and politically motivated and accusing Washington of attempting to suppress competitors under the banner of intellectual-property protection and national security.

Yet Beijing’s rejection of American threat narratives should not be interpreted as a belief that advanced AI is harmless.

China itself is preparing for the possibility that future systems could become difficult to control.

Chinese State Security Minister Chen Yixin warned in an article published Sunday that increasingly powerful foreign AI models could threaten China’s critical information infrastructure and national security. Chinese regulatory frameworks have also begun addressing scenarios involving autonomous systems that can replicate, evolve or operate beyond expected human supervision. President Xi Jinping has emphasized that artificial intelligence must ultimately remain under human control.

That creates an unusual situation: the United States and China increasingly agree that powerful AI may generate serious risks, but they disagree deeply over where those risks come from and who should be trusted to manage them.

Washington worries that advanced Chinese AI could strengthen Beijing’s military, surveillance and cyber capabilities.

Beijing worries that powerful American closed models could create security dependencies and vulnerabilities inside Chinese networks.

At the same time, both fear falling behind the other.

The competition is increasingly moving beyond technology itself and into the architecture of global AI governance.

China has been building its own diplomatic framework for artificial intelligence, particularly among emerging and developing countries. On July 16, representatives from 29 countries signed an agreement establishing the World AI Cooperation Organization, an intergovernmental body backed by Beijing and intended to promote international cooperation and AI governance.

China has presented itself as a champion of affordable and broadly accessible AI, arguing that the technology should not become the exclusive property of a handful of wealthy Western nations or corporations.

That effort accelerated at the BRICS summit in New Delhi on September 13.

President Xi Jinping announced plans for China to lead the creation of a BRICS AI Open Source Zone, designed to encourage cooperation on large language models, AI training and an open artificial-intelligence ecosystem among BRICS countries. The initiative fits Beijing’s broader effort to expand its technological influence across the Global South.

The battle is therefore becoming much larger than OpenAI versus DeepSeek or Anthropic versus Alibaba.

The emerging contest is over the technological foundations of the next industrial era: who supplies the chips, whose models countries use, where data is processed, which standards governments adopt and whose vision of artificial intelligence becomes embedded in national infrastructure.

Financial markets are already responding to the growing uncertainty.

AI-related stocks came under pressure on September 14 after the industry’s most prominent executives called for slower capability development. Investors began questioning whether the enormous capital spending planned for AI chips, data centers and energy infrastructure could be affected if companies voluntarily or regulatorily limit the pace of frontier-model expansion.

The next major test could come remarkably soon.

President Xi is scheduled to visit the United States on September 24, when he is expected to meet President Trump. Trump has already said artificial intelligence will be discussed during the visit. The two leaders also discussed AI during Trump’s earlier trip to Beijing.

Washington and Beijing have also been preparing discussions focused specifically on frontier AI safety risks, making the coming diplomatic engagements potentially significant for the future of global technology governance.

But creating anything resembling an international arms-control regime for AI will be extraordinarily complicated.

Nuclear weapons can be counted. Missile silos can be photographed. Warships can be monitored by satellites.

AI models are different.

Their most important ingredients—algorithms, data, model weights, software and computing techniques—can be copied, concealed and transferred far more easily. A breakthrough developed inside one laboratory could spread internationally within weeks or even days.

Open-weight systems create an additional dilemma. They allow researchers and cybersecurity specialists to inspect and improve models, but once powerful model weights are publicly distributed, preventing modification or malicious use becomes far more difficult.

This technological reality produces a dangerous strategic incentive.

If Washington believes Beijing is racing ahead, American laboratories face pressure to accelerate.

If Beijing believes Washington is racing ahead, Chinese developers face exactly the same pressure.

Neither side wants to be the country that behaved cautiously while its rival achieved a decisive technological advantage.

That is why the deepest challenge facing AI governance may not be whether scientists understand the dangers. Increasingly, many of them do.

The challenge is whether governments and companies can act on those concerns while trapped inside an international competition that rewards speed.

There are still areas where cooperation may be possible. Washington and Beijing could develop minimum safety standards for the most capable systems, establish communication channels for major AI-related cyber incidents, cooperate on preventing AI assistance in biological-weapons development and create testing requirements for models capable of operating autonomously against critical infrastructure.

They could also establish rules for military AI, where autonomous decision-making carries potentially enormous consequences.

Such agreements would not require either country to surrender its technology or abandon competition. They would instead establish boundaries designed to prevent the rivalry from producing an uncontrolled technological escalation.

But achieving even that limited cooperation will require something currently in short supply: trust.

The global AI race has therefore reached a pivotal point.

For Silicon Valley, the question is how rapidly increasingly capable systems should be built.

For Washington, the question is how to preserve American technological leadership without creating unacceptable security risks.

For Beijing, it is how to challenge American dominance while ensuring China is not locked out of the global rules governing the technology.

For the rest of the world, however, the stakes are even larger.

Artificial intelligence is becoming infrastructure for economies, governments, defense systems and scientific research. Decisions made today in Washington, Beijing and a handful of technology laboratories could shape international power for decades.

The defining question of the AI age may therefore no longer be simply whether America or China wins the race.

It may be whether the world’s two most powerful AI nations can compete intensely enough to innovate—yet cooperate enough to prevent that competition from becoming a race neither side can safely control.