Despite a 17% Plunge, Why Will NVIDIA Benefit from DeepSeek?

Introduction

This year's Spring Festival was not a joyful occasion for investors in U.S. stocks, especially those holding NVIDIA shares. Before the festival, as the DeepSeek inference model R1 gained popularity, NVIDIA's stock price suffered a sudden and significant drop, plunging 17% in a single day and losing approximately $600 billion in market value. In the following days, the stock continued to decline, reaching a new low in recent months.

Some investors attributed this decline to the fact that DeepSeek's training cost was extremely low, at only $5.576 million, yet its performance was close to that of OpenAI's o1 model. This achievement broke the industry's long-held belief that "high performance equals high computational investment," leading the market to worry that DeepSeek's low-cost training method might reduce the demand for NVIDIA's high-end products such as the H100 and B200, thereby impacting NVIDIA's business and its high-profit margins.

However, during the Spring Festival, I joked in the training camp community that this was actually a disguised gift from DeepSeek, offering a discount on NVIDIA's stock. Subsequently, as the market gained a deeper understanding of DeepSeek-R1, coupled with NVIDIA's strong financial performance and the capital expenditure plans disclosed by major tech companies, NVIDIA's stock price gradually began to recover.

It is still important to note that in the short term, the capital market is always full of uncertainties and often overreacts to information. Therefore, investment should be approached with caution. But I believe that due to the emergence of DeepSeek, NVIDIA's performance this year will be better than expected and is likely to be reflected in its stock price.

Why NVIDIA Will Benefit from DeepSeek

1. Increased Demand for Computing Power

Although DeepSeek's low-cost training has indeed reduced the consumption of high-end GPUs for single model development, it will lead to more AI applications participating in training and create a broader demand for computing power in the inference stage. The R1 model of DeepSeek marks a rapid reduction in the cost of large models, which will help popularize AI technology and enable more enterprises and individuals to access advanced AI technology. As a result, the demand for both training and inference will experience explosive growth.

This growth in demand will directly drive the need for computing infrastructure. For example, tech giants such as Google, Meta, and Amazon have all indicated that they will continue to increase their capital expenditure on cloud computing infrastructure. NVIDIA has also responded by stating that DeepSeek will bring more demand for their products and pointed out that the computing power demand in the inference stage still highly depends on NVIDIA's hardware.

2. Expansion of the Open-Source Ecosystem

The open-source ecosystem driven by DeepSeek is continuously expanding, which may encourage more enterprises to participate in AI application deployment. With the proliferation of open-source AI models, third-party AI service providers will emerge, offering customized AI solutions for enterprises. For instance, companies like Oracle and IBM can attract a broader customer base by providing platforms that integrate open-source AI models. This will expand the overall application volume of AI, thereby driving the demand for computing power.

3. Challenges and Opportunities in the Inference Stage

For the inference stage, dedicated architecture chips such as Broadcom's XPU and other ASIC chips (Application-Specific Integrated Circuits) may have a stronger cost-performance advantage compared to general-purpose GPUs. This could potentially pose a challenge to NVIDIA.

Of course, this impact does exist, so NVIDIA is not without worries. It faces a situation where challenges and opportunities coexist. On one hand, NVIDIA will face the challenge of a market structure shift. In the past, its growth mainly relied on the concentrated procurement of high-end GPUs by leading tech companies for AI model training. However, with the rise of the open-source ecosystem, the demand for computing power is dispersing from the training stage to the inference stage, which will require adjustments to NVIDIA's product line.

Customers focusing on inference will be more inclined to purchase NVIDIA's mid-to-low-end hardware. For example, traditional hardware vendors like Oracle, once they provide inference services based on open-source models, may shift to more cost-effective product lines instead of the expensive B200 and other top-tier GPUs.

On the other hand, NVIDIA is expected to dominate a significant portion of the新增 (newly added) inference market. We have observed that as open-source models also achieve high performance, third-party AI service providers will join in to offer more personalized and customized services for users. They can rely on more affordable and fragmented computing power, i.e., the so-called "small clouds," for inference, or even purchase chips to build their own servers, instead of solely relying on the support of large cloud platforms.

For these "small clouds," purchasing NVIDIA's GPUs remains the most economical and convenient choice. Although ASIC chips like Broadcom's XPU are not controlled by NVIDIA and can offer faster inference speeds, their development costs are high. Only "large clouds" like Amazon, Google, and Microsoft, which require hundreds of thousands of chips, would consider using ASIC chips to reduce costs and improve efficiency. It's worth noting that even Oracle has not engaged in ASIC chip development but instead purchases NVIDIA's GPUs.

Moreover, NVIDIA has launched the NVIDIA NIM inference microservices platform, which can deepen its presence in the enterprise market by bundling hardware and software services, thereby increasing customer loyalty.

Conclusion

In conclusion, although DeepSeek has attracted a lot of praise and attention through its low-cost, high-performance AI solutions, it is important to note that the Scaling Law has not become ineffective. In the long term, continuously increasing investment and computing power remains the main way to enhance the performance of AI models.

As we have repeatedly emphasized, NVIDIA has built a strong technological moat in the data center field with its ecosystem, CUDA ecosystem, and hardware iteration speed, and will still occupy a leading position in the global market in the future.

Therefore, the real threats to NVIDIA may only come from the U.S. government's export restrictions and various reviews of NVIDIA. Of course, I also want to say that in the face of the AI revolution, it seems a bit of a pity to only focus on the investment level. Transforming our careers, industries, and even solving common pain points in work and life through AI is what should be done in the AI revolution.