NVIDIA CEO Jensen Huang said in an interview with the Dwarkesh Patel podcast on Wednesday, April 15, that the type of compute required to train the Anthropic AI model Claude Mythos is “very plentiful in China,” and that existing data center infrastructure in China is sufficient to support training models of comparable scale.
Jensen Huang’s statement: The current state of China’s compute and how to respond to U.S.-China competition
According to the content of Dwarkesh Patel’s April 15 podcast interview, Huang said that for Claude Mythos’s compute needs, the model is trained on “fairly standard training,” and the type of compute required is “very plentiful” in China.
Huang quoted in the interview: “Their data centers are completely idle, yet their power supply has not diminished in the slightest. Their infrastructure capacity is so enormous. If they wanted, they could build some additional chips.”
Huang also noted that China produces 60% of the world’s mainstream chips, has 50% of the world’s AI research personnel, and has abundant energy resources. Regarding U.S.-China AI competition strategy, Huang said in the interview: “Treat them as victims, turn them into enemies—maybe that’s not the best way to solve it. They are the opponent.” And “We want the United States to win. But I think engaging in a dialogue—especially a research-oriented dialogue—may be the most robust approach.”
Claude Mythos security assessment: Reports from AISI, Anthropic, and Reuters
According to an assessment report released by the AI Safety Institute (AISI) on April 13, Claude Mythos can carry out multi-stage attacks on vulnerable networks, independently discover and exploit vulnerabilities, and the corresponding tasks would require human experts to spend several days to complete.
The test results for the Claude Mythos Preview released by Anthropic on April 7 showed that the model can find zero-day vulnerabilities, and that 99% of the vulnerabilities it discovered had not yet been fixed. After that, Anthropic restricted access to Claude Mythos.
According to a Reuters report on Tuesday, April 14, AI-enhanced hacking attacks could affect banking systems that use software from decades ago.
Background: Statements by the Treasury Secretary and historical reporting by Anthropic
According to a report by Bloomberg on Tuesday, April 14, U.S. Treasury Secretary Scott Bessent characterized Claude Mythos as “a revolutionary advancement in AI,” and said the model will help the United States stay ahead in the AI race. Bessent was quoted as saying: “Mythos achieves a qualitative leap in capability and learning ability.”
According to a report from Anthropic last November, an “organization supported by the Chinese government” manipulated its Claude Code tool to attempt to infiltrate around 30 global targets, and succeeded in a small number of cases.
FAQ
Where and when did Jensen Huang make his statement about China’s AI compute?
According to the Dwarkesh Patel podcast, Huang made the above comments regarding China’s compute and the training needs for Claude Mythos during his April 15 (Wednesday) interview with Dwarkesh Patel. The podcast is hosted by Dwarkesh Patel.
What conclusions did AISI reach in its assessment of Claude Mythos, and when was it published?
According to an assessment report released by the AI Safety Institute (AISI) on April 13, Claude Mythos can carry out multi-stage attacks on vulnerable networks, independently discover and exploit vulnerabilities, and complete tasks that would otherwise require human experts several days to finish.
When did Anthropic restrict access to Claude Mythos, and why?
According to Anthropic’s announcement, the test results for the Claude Mythos Preview were released on April 7, showing that the model can find zero-day vulnerabilities and that 99% had not yet been fixed. After that, Anthropic restricted access to the model.
