Note: the title of this article is generated by DeepSeek R1. The article itself, though, is written by a real human.

Chinese gen AI startup DeepSeek has suddenly and forcefully disrupted some of the conventional thinking in generative AI.

Overnight (27 Jan), the Nasdaq index fell more than 3%, wiping out trillions of dollars of market value—Nvidia alone lost US$593 billion. The widely cited reason is that investors worry that DeepSeek has demonstrated that US chip and tech giants’ dominance in global AI development might not be as solid as once thought.

Over the last weekend, we have spoken intensively with some close friends who are researchers and CTOs of AI companies across a couple of cities in China. We are sharing the summarized notes of some perspectives below.

You probably have read a lot about DeepSeek over the last couple of days—we will not bore you with something you already know or can read elsewhere.

Here you go:

  1. DeepSeek is the first LLM chatbot our friends find usable on a daily basis: it is fast and demonstrates the full reasoning process. In comparison, ChatGPT is slow, hallucinating, poor in real-time online search, and often too conservative in its response. And of course, when accessing ChatGPT from China, you have to worry about the stability of VPNs.
  2. That said, experience shows that GPT 4 overall still has better logic. Also, for business use, the ChatGPT API is still the best in the market: more stable, albeit very expensive compared to DeepSeek.
  3. One friend predicted a year ago that some non-state force in China would eventually figure out what DeepSeek has achieved—training top models at low costs. However, he thought the breakthrough would come from crypto miners, not a quant fund.
  4. Alexandr Wang, CEO of Scale.ai, claimed in a CNBC interview that DeepSeek had 50,000 H100 GPUs. That is unlikely but not impossible—DeepSeek has an incentive not to reveal how many GPUs it really has because there is still a long way to go.
  5. Venture investors should probably give up investing in new LLM companies if they want to stick to the valuation models they are familiar with. Someone will win eventually—we do not know who, or when. Until then, it would be an incessant race with continuous investments needed and big tech/state players involved.
  6. DeepSeek does not have to worry about this because it already has tens of thousands of GPUs. It will eventually have to work out how to make money but does not need to worry now since it is not fighting for user numbers or other valuation metrics mentioned in point 5.
  7. There is enough engineering, operations, and business talent in China to allow DeepSeek to continuously build on. A related point is that DeepSeek’s team employs young engineers rather than big shots—its star engineer (a 29-year-old woman) was poached by Xiaomi with a >4X salary increase, causing no impact on DeepSeek’s ability to carry on.
  8. It is unlikely that DeepSeek will cause deflation of demand for NVIDIA and others’ chips. While DeepSeek has plausibly proven that the demand for training compute should not be as high as previously thought, it also shows reasoning that might end up consuming a lot more computing power.
  9. DeepSeek leadership knows exactly how to distribute news to elicit a market response. Last night’s release of its multimodal model Janus-Pro-7B is a testament to their ability to time market reactions effectively.
  10. The competition between China and the US in AI is now in the open—for the better. President Trump’s strategy of generating more electricity and lowering energy costs might provide a long-term advantage in this open AI competition.
  11. You should really send your kids to Zhejiang University, where DeepSeek’s founder Liang Wenfeng studied. The University is also the alma mater of other top Chinese business figures like Duan Yongping and Pinduoduo founder Colin Huang. There must be something they are doing right.