Reasons for Nvidia's plunge and outlook

Source: Dong Zhidao Research

When Lei Jun, wearing a brown leather jacket, released new products and became the industry's "price butcher", Huang Renxun, wearing a black leather jacket, once again faced a sharp drop in Nvidia's stock price because of the "price butcher" in the industry: the big negative line of 8% again aroused market concerns.

Nvidia's current round of stock price decline began withDeepSeek: it achieved the same model performance with only1/10 the computing power cost of its peers. A big model efficiency revolution has broken the convention of large computing power, and has also caused investors to think: the computing power required to train models may be much lower than expected. A few days ago, Huang Renxun publicly responded to the impact of DeepSeek for the first time when he attended a DNN event. There are three general opinions: 1. DeepSeek is truly amazing. (People who don't believe that domestic large-scale models are really amazing should be able to believe it now.) 2. Investors have a completely wrong understanding. “From an investor’s perspective, they think the world is divided into two stages: pre-training and inference, and inference is asking questions to the AI ​​and getting answers immediately. I don’t know who created this misunderstanding, but it’s obviously wrong.” 3, post-training is still important and requires a lot of resources; and inference itself is the “computationally intensive part.” There is nothing wrong with these views. However, what Huang did not point out is that for investors, the "main contradiction and the main aspects of the contradiction" have changed. Before the release of DeepSeek-R1, the market's focus was on pre-training and NVIDIA. However, after the release of R1, the market's focus switched to reasoning and low-cost computing power.

In the training phase, the high throughput and high parallel speed of NVIDIA chips ensure that the company is far ahead; however, in the inference phase, especially the innovative solutions adopted by R1, the threshold of the required chips has been lowered.

In other words, it is not that NVIDIA products are no longer good, but that other companies' products can also be used.

This kind of contradictory switch will inevitably cause investors to rethink. Even if the company's current and future performance is still excellent.

From this point of view, investors in US and A-share technology stocks are the same. Technology stocks are almost never driven by valuations. No company has soared because of low valuations, and no company has plummeted because of high valuations.

The core drivers are: marginal changes and changes in contradictions.

So, what will happen to Nvidia in the future?

In the medium term, with the expectation of reasoning, it will be gradually recognized by the market; if large models really begin to be widely popularized and penetrate the industry, it will inevitably drive the demand for training again. The market contradictions are cut back to training.

In the long term, with every passing day, Nvidia's risk is getting closer: Huang Renxun retires.

The importance of Huang Renxun's personal temperament to NVIDIA is self-evident. The company's management culture and innovation mechanism are inseparable from Huang Renxun's day and night polishing. If one day, Huang retires and can no longer manage the company, the successor at that time is likely to be inferior to Huang Renxun.

The company's manager rule is: the first generation has the strongest ability and opens the company's glory; then there are the key entrepreneurial ministers, who are also capable and can maintain the business and develop; then there are either key ministers of key ministers, or professional managers.

Choosing a professional manager is like opening a lottery and hitting the jackpot, with mixed results.

Microsoft missed the mobile Internet during Ballmer's time; under Nadella's leadership, it caught the wave of cloud computing; Intel, under Grove's "Paranoia is the only way to survive" philosophy, created a chip empire, but then missed the wave under the leadership of professional manager Otellini, and the next twoCEOsalso hesitated and now has become an acquisition target.

Therefore, Nvidia may not be able to escape this situation.

Another point is the growth of Chinese competitors. Compared with domestic chip companies, NVIDIA has two competitive advantages: manufacturing technology and CUDA ecosystem. The former depends on the execution of the national chip strategy, whether SMIC can withstand the pressure and continue to invest; whether equipment vendors can achieve the effect of DeepSeek with existing resources, or directly break through the technology. As long as you spend money and time, you will have a chance. It will take about ten years for CUDA ecosystem to be developed. If college students learn various courses based on domestic GPUs during college, and there are about four or five graduates, it will be a good start. Back to the current NVIDIA, the company has another important advantage, which is Huang Renxun and NVIDIA's learning and error correction ability (derived from the diligence of managers, the company's high talent density, and the system and culture of believing in talents, motivating talents and rewarding talents). From a historical perspective, when NVIDIA is in crisis, it is often more cost-effective. Once I asked a friend who works at NVIDIA how to look at the company's stock price and whether there is any swing operation. His answer was that there is very little swing operation, first, it is not accurate, and second, it is meaningless. As long as the company's colleagues are extremely diligent in innovation, the stock price will definitely rise again.

After all, from a broad perspective, AIhas just begun to enter our lives; there are still many problems that can be solved by GPUparallel computing.