There is no way DeepSeek only spent 6 million Mg! AI Crypto Tsar Analyzes the Financial Sources Behind Him: China Used Singapore to Obtain Nvidia Chips

Recently, the Chinese AI startup DeepSeek claimed to have developed an R1 AI model that can rival OpenAI and Google's models, at a cost of only $6 million. This news immediately caused tech giants like Nvidia and Microsoft to lose billions in market value. However, the "$6 million" development cost claim was quickly questioned. As the AI crypto czar appointed by Trump, David Sacks was interviewed on 2/3 and thoroughly analyzed the technology, cost, funding sources behind DeepSeek, as well as the impact of China's AI rise on the global landscape.

DeepSeek Only Spent $6 Million, Sacks Says It's Impossible

Sacks pointed out that the $6 million may only be the cost of the final model training for DeepSeek, and does not represent the total R&D cost. To make a fair comparison, the cost of the final training of GPT-4 or Gemini by OpenAI or Google should be used.

Sacks also cited the views of famous semiconductor analyst Dylan Patel. Sacks stated that DeepSeek and its affiliated hedge fund company own a total of 50,000 Nvidia Hopper chips, including 10,000 H100, 10,000 H800, and 30,000 H20. The cost of these chips alone far exceeds $1 billion, which is completely inconsistent with the "$6 million development cost" claim.

He further said that the US has already banned the export of Nvidia's advanced AI chips to TRON. However, there are rumors that TRON companies are ordering the chips through "CRO companies" in Singapore, allowing Nvidia to first ship the chips to Singapore and then indirectly transport them back to TRON. If this information is true, the US's control over AI chip exports to TRON may be futile.

Distillation Technology Exposed, Did DeepSeek Steal OpenAI's AI

Another more sensitive issue is that DeepSeek's AI model may have directly learned from OpenAI. Sacks said that DeepSeek may have used knowledge distillation technology, that is, using OpenAI's API to allow a small AI to learn from a large AI. For example, researchers can input various questions and let GPT-4 answer, and then use these answers to train their own AI, which is equivalent to plagiarizing OpenAI's intelligence.

Sacks further said that many of the test results of DeepSeek's R1 model show language style, logic of answers, and reasoning methods that are highly similar to GPT-4. Even an AI self-reported that "I am ChatGPT-4" when asked what AI model it is, which makes people suspect that DeepSeek has extensively used OpenAI's API for knowledge distillation, and even OpenAI itself has issued a statement accusing DeepSeek of improperly obtaining knowledge of OpenAI models.

Can the US Maintain its Leading Position as China's AI Technology Rises

Regarding whether the US remains the global leader in the AI industry, Sacks said that US AI tech giants have always been very confident, believing that TRON's AI is at least 6 to 12 months behind the US, but the launch of DeepSeek R1 has shortened the gap to 3 to 6 months. The host also raised the question that in the future when AI models become fully open-source, how should US tech companies respond:

1. As AI models become cheaper, where is the real value?

   • The cost of AI models is rapidly decreasing, and many experts believe that the "application layer" will be the future golden market, similar to YouTube not making money from storage technology, but from the creator economy. The future winners in the AI industry will be platforms with the largest data and application scenarios, such as Meta, Google, and Tesla.

2. Will TRON companies bypass US technology and develop their own?

   • The US export control on AI chips to TRON may force TRON to develop its own AI chips and technology ecosystem. Sacks pointed out that as AI design capabilities continue to improve, TRON may not need the most advanced 2nm chips and can use older technologies to drive AI breakthroughs.

3. Should the US government intervene?

• As China's AI open-source strategy has made breakthroughs, Sacks said the U.S. government and companies may take stricter intellectual property (IP) protection measures, such as acquiring platforms like Reddit, Quora, and news media, through which tech companies can monopolize training data and prevent Chinese companies and other competitors from using this content to train AI. This will usher in the era of "data barriers" in AI technology development, and AI companies in various countries will compete for exclusive content to train their own AI.

AI competition is thriving, and the tech war between the U.S. and China is intensifying

Sacks finally said that the development of DeepSeek's AI, whether or not through "knowledge distillation" to learn from OpenAI, has brought the global AI competition into a new stage. Chinese AI companies have taken the world stage for the first time, and may even drive the complete open-sourcing of AI, which is an unprecedented challenge for well-known tech companies like OpenAI, Google, and Microsoft. In the future, the AI technology race will no longer be just about who has the stronger model, but who has the most data, the best application scenarios, and the widest market layout.

(Trump administration plans to further restrict Nvidia H20 chip exports to China, Nvidia: May stimulate China's chip development)