China Tech Has Entered Bull Market
Fri Jan 31 2025
Dear readers,
In my previous newsletter, I wrote on the China put and the healthy part of China. For discerning investors, the return in the healthy part of China has definitely outperformed the global market in absolute terms. What was the key driver? China AI.
I have used many other AI tools before, such as OpenAI’s ChatGPT and Google’s Gemini, which directly give a result, and the output result is amazing but also what is expected. Categorically, the results still fall under a probabilistic outcome within data simulation, and is not a replacement for our reasoning process that has been refined through experience — the ability to reason remains our domain and triumph of our humanity.
Until I saw DeepSeek list the reasoning process bit by bit, I couldn't tell whether it was a human or a machine behind it. At that moment, a feeling of the excitement came to my mind, this made me rethink the impact and opportunities that Deepseek's emergence has on the AI industry, the world, China, and ADDX when we are still in the beginning of AI revolution.
Significance of emergence of DeepSeek R1
Innovation is always the primary productive force, the most eye-catching thing about DeepSeek is that it proves pure outcome reward RL(reinforce learning) can directly raise the model to the ChatGPTo1 level. Before it came out, everyone in the industry (including DeepMind) believed that PRM (Process Reward Model) was needed to achieve this. This is already a discovery that subverts the industry’s assumptions. Now all LLM groups except GPT are starting over and copying DeepSeek training methods. In addition, it is very important that DeepSeek has also discovered that this training method can even allow the model to learn longer-chain reasoning and reflection by itself, which they call the "aha moment". It is equivalent to only training LLM to get more accurate results. LLM can learn to self-reflect and know that it is on the wrong track of reasoning, and then try to course-correct the error by itself. The "self-evolution" feature of this model is a major discovery in the industry, second only to GPT intelligence emergence.
Why did it cause such a huge reaction on both AI industry and capital market?
The current AI landscape is dominated by models from tech giants. DeepSeek is game-changer for AI accessibility in the following angles:
To AI app builder: DeepSeek's disruptive pricing strategy - DeepSeek R1's API is 27 times cheaper than OpenAI's GPT models while delivering comparable performance, it probably pushes the entire industry toward more affordable AI solutions. Many innovative and practical AI applications will be implemented at lower costs and fundamentally change the efficiency and competitiveness of individuals and industries.
To model provider: "training a model with similar effects with fewer and low-profiling GPU" may not only save costs, but also be an improvement of scaling law, which means that this method of stacking more cards may increase the model's capabilities by an order of magnitude, or even directly reach AGI/ASI. This is why the industry is so hyper this time. The value of DeepSeek open source is far greater than Meta’s LlaMA. LlaMA is basically a known method of stacking cards for training.
To SMEs: The open source of DeepSeek has greatly lowered the entry threshold for top AI models. Previously, IT departments in various industries were unable to develop their own customized AI models fine-tuned with exclusive data and could only rely on giants to provide interface services at high prices. Now, relying on the open source DeepSeek big models, they can reduce their dependence on AI giants.
How it will massively influence the tech development and economic development in China and globally
DeepSeek will be good for almost all AI chains from applications, power, to ASIC chip designer but also bad news for Nvidia in the short and mid-term, but probably good news in the long-term. Because in the short term, the business model of Meta, Google, Microsoft, etc., which purchase many Nvidia chips and charge by tokens, has been weakened. Many corporate IT departments were originally unable to develop their own low-cost top-level large models and had to rely on giants or third-party professional AI companies for outsourcing or API. Now they can build their own top models with open source like DeepSeek or low-cost API. The barriers of AI giants have dropped, and the expected return on investment has dropped, which will slow down the scale and pace of buying GPUs, and Nvidia's short-term performance and price premium will be affected.
But in the long run, the technological progress and open source of DeepSeek and other models will promote the development of the industrial chain. The previously unfeasible AI commercial applications have suddenly become economically feasible. The explosion of AI applications will eventually lead to the explosion of AI terminals (such as AI mobile phones, AI computers, AI glasses, and various AI robots), which will accelerate the demand for AI chips on the consumer side and small businesses in the long run.
For China, we are seeing more top model with same MoE(Mixture of Expert) architecture emerging i.e Alibaba's latest launch of Qwen2.5-max which also par with GPTo1 performance, the overall token price of China model is much cheaper than equivalent ones in us, therefore China will emerge many killer and popular AI applications in 2025. We can pay more attention to such opportunities in investment and primary market. And as we assume more restrictions on GPU exports from the United States, China's self-owned GPUs such as Huawei Ascend and Cambrian will develop and adapt to China's top large models in faster pace since the innovation of DeepSeek model algorithm and architecture has gotten rid of the necessity of high-end GPU for top-level large model training and inferencing.
Should you chase the AI rally? I personally think we are only at the beginning of the AI cycle in China/Ex-US.
Chairman