DeepSeek claimed to have done what U.S. AI companies did, but with far less training (on specialized AI semiconductors) and for a fraction of the cost. This created a panic among AI investors. If DeepSeek can do AI without much computing power (known as “compute”), then the companies that power AI with chips, energy, and infrastructure, goes the reasoning, are overvalued.
But later reports suggest that DeepSeek may have exaggerated the efficiency of its AI training regime. OpenAI and the Trump administration’s AI adviser raised questions about whether DeepSeek inappropriately accessed OpenAI content, with which it “distilled” its LLM. And Washington is probing whether DeepSeek bypassed U.S. export controls on AI chips by purchasing them through Singapore.In 2023, High Flyer nevertheless started focusing on artificial general intelligence (AGI), a milestone when AI surpasses human intelligence for most economic activities. According to a report on Jan. 6, DeepSeek has far more computing than previously acknowledged. The report cites a source indicating that DeepSeek has 50,000 H100 chips, far more than it publicly acknowledges. DeepSeek allegedly did not disclose these chips, each of which is about three times more powerful than an A100, because of U.S. export controls. But even 50,000 H100s are far less than some of its competitors. For comparison, Meta uses the equivalent of 600,000 H100 chips.
DeepSeek has other problems. It reportedly includes secret code that sends personal data back to China, where, by Chinese law, the regime can use it for anything it likes. And DeepSeek is loath to say anything critical of the Chinese Communist Party (CCP). So the more DeepSeek proliferates, the more the CCP gets a free pass.
DeepSeek has minimal safety features, according to Anthropic and Cisco security researchers. Their tests indicate that DeepSeek is particularly vulnerable to “jailbreaks,” user manipulation of prompts to make AI divulge dangerous information. Just wait until TikTok lets users deploy DeepSeek to create their posts. Then we’ll all be in big trouble.
Given these and other risks from AI, the reasonable response by regulators and legislators around the world would be to ban DeepSeek. Yet only Italy has fully banned the app. DeepSeek did not adequately respond to the questions posed by Italian regulators regarding the types of personal data collected from users and whether that data is stored in China. DeepSeek reportedly said it should not be subject to Italy’s jurisdiction, local regulators, or obligations to provide information. This worsened DeepSeek’s position with respect to Italy, which is what prompted the ban. Texas, the Pentagon, the U.S. Navy, and NASA have banned the app on their devices. U.S. legislation is in the works to ban it, but only on government-owned devices.India, Australia, and South Korea have already banned the app on government devices but still allow downloads to commercial and private devices. France and Ireland are also probing DeepSeek on the issue of privacy but have not banned it yet.
Given the difficulty of banning TikTok after it was allowed to grow in popularity for too long, U.S. lawmakers should ban not only DeepSeek but also all apps from China and other adversary nations. We can no longer have a presumption of innocence when it comes to technology under the control of the CCP and other malign actors. Whack-a-mole does not work, especially when the moles first grow into behemoths.
