Nvidia CEO Jensen Huang has warned that China possesses the massive computing power and infrastructure necessary to match the capabilities of Anthropic’s high-level AI model, Claude Mythos.
Summary
- Nvidia CEO Jensen Huang warned that China’s vast infrastructure and “ghost datacenters” provide the necessary computing power to match the capabilities of Anthropic’s Claude Mythos.
- Huang stated that the hardware used to train the revolutionary model is abundantly available in China and suggested the Chinese government could easily mobilize more chips to rival American AI developments.
Speaking on the Dwarkesh Patel podcast on Wednesday, Huang dismissed the idea that China lacks the hardware to keep pace with Western AI breakthroughs.
While Washington views the recently released Mythos model as a critical advantage, Huang noted that the capacity required to train such a system is “abundantly available” in China.
He pointed to a vast landscape of “ghost datacenters” that sit empty but fully powered, suggesting that the Chinese government could “gang up more chips” at any moment to rival American developments.
The urgency of this parity stems from the specific cyberoffensive capabilities found within the Anthropic model. Internal testing in April revealed that Mythos could identify thousands of software vulnerabilities in browsers and operating systems, prompting the company to limit access to prevent misuse. The AI Security Institute backed these concerns on April 13, finding the model capable of executing autonomous, multi-stage attacks on networks.
Financial institutions remain particularly vulnerable to these AI-boosted hacks because they rely on decades-old software, according to a Reuters report published Tuesday.
These fears are grounded in recent history; Anthropic reported in November that state-sponsored actors from China had already attempted to use the “Claude Code” tool to infiltrate dozens of global targets.
Balancing competition with dialogue
U.S. Treasury Secretary Scott Bessent has described Mythos as a “step function change” in learning capabilities that would secure American leadership.
Huang, however, argued that treating China solely as a target for restrictions could be counterproductive, given that the country produces 60% of the world’s mainstream chips and hosts half of all AI researchers.
“Victimizing them, turning them into an enemy, likely isn’t the best answer,” Huang said.
Though he identified China as an adversary and expressed a desire for the United States to win the technological race, he maintained that the sheer scale of Chinese infrastructure necessitates a more nuanced approach.
“I think having a dialogue and having research dialogue is probably the safest thing to do,” he added, suggesting that open communication might be the only way to mitigate the global risks posed by such powerful autonomous tools.
Last week, Bessent met with executives at a major American bank last week to address what he described as “unprecedented” cybersecurity risks posed by the next generation of autonomous AI.
