NVIDIA Drops Nearly 10% in Pre-market Trading! "Sense of Fear" Grips Silicon Valley! Does This Chinese-listed Stock Have a Close Ties with DeepSeek?

On January 20th, Hangzhou DeepSeek, a Chinese large-language-model company, officially released the DeepSeek R1 model.

 

This model showcases performance in key areas such as mathematics, programming, and reasoning that can even rival OpenAI's most powerful reasoning model, o1. However, its API call cost is 90%-95% lower.

 

In just one week, the superior performance and ultra-low cost of this latest model have sent shockwaves through Silicon Valley, easily creating a sensation in the global AI community.

 

DeepSeek didn't "amaze" everyone out of the blue.

 

 

At the end of 2024, DeepSeek released the new-generation MoE model, DeepSeek-V3. It has 671 billion parameters, with 37 billion activated parameters, and was pre-trained on 1.481 trillion tokens.

 

In knowledge-based tasks (MMLU, MMLU-Pro, GPQA, SimpleQA), V3 is close to the currently top-performing Claude-3.5-Sonnet-1022. In terms of code-writing ability, it is slightly better than the latter. In terms of mathematical ability, V3 has clearly outperformed other open-source and closed-source models, including Llama3.1 405B-Inst, GPT-4o 0513, and Qwen2.5 72B-Inst.

 

This is already an excellent open-source model. However, what really drew a large amount of attention was that in the technical paper, DeepSeek stated that the total training cost of the DeepSeek-V3 model was $5.576 million, and the complete training consumed 2.788 million GPU hours, almost one-tenth of what is required for models of the same performance level.

 

This is one of the core performances that attracted Meta's attention to DeepSeek-V3. What pushed this level of attention to a new height was the reasoning model R1 released by DeepSeek a week ago.

 

On January 20th, DeepSeek released DeepSeek-R1, which is performance-aligned with OpenAI-o1 official version, and simultaneously open-sourced the model weights.

 

It is basically on a par with OpenAI-o1-1217 in tasks such as mathematics, code, and natural-language reasoning, especially winning by a narrow margin in the three test sets of AIME 2024 (American Invitational Mathematics Examination), MATH-500, and SWE-Bench Verified (a test set in the software development field).

 

As a verification of R1's capabilities, among the multiple small-sized models distilled from the 660B version of R1, the 32B and 70B models can be benchmarked against OpenAI o1-mini in multiple capabilities.

 

Moreover, these distilled models belong to the Qwen series and Llama series. Among them, the 14B Qwen-series distilled model has significantly better performance in various reasoning-type test sets than QwQ-32B-Preview.

 

What was more eye-catching at that time was the simultaneous open-sourcing of DeepSeek-R1-Zero, a result obtained by adding only RL (Reinforcement Learning) on the basis of pre-training without going through SFT (Supervised Fine-Tuning).

 

Due to the lack of human-supervised data intervention, R1-Zero may have problems such as poor readability and language mixing in generation. However, this model is still comparable to OpenAI-o1-0912.

 

Its more important significance is that it explores the technical possibility of obtaining reasoning ability by training large-language models only through reinforcement learning, providing an important basis for subsequent related research.

 

As discussions about DeepSeek models providing high cost-effectiveness with low-capacity chips continue to increase, this may raise questions about the dominance of US technology companies, including NVIDIA.

 

Nirgunan Tiruchelvam, the head of consumer and internet at Singapore-based Aletheia Capital, said that there was a claim that the huge capital expenditures and operating expenses in Silicon Valley were the most appropriate way to approach the AI trend.

 

However, the emergence of DeepSeek's products has dealt a huge blow to this claim, raising doubts about the large amount of resources invested in the AI field.

 

As US futures declined, major technology companies such as Apple and Microsoft are about to release key financial reports this week.

 

 

Charu Chanana, the chief investment strategist of Saxo Bank, said, "Although current leaders like NVIDIA have a firm foothold, this reminds us that the dominance in AI cannot be taken for granted... The emergence of DeepSeek shows that competition is intensifying. Although it may not pose a significant threat now, future competitors will develop faster and challenge existing companies more quickly. This week's financial reports will be a huge test."

 

It's not difficult to see that the success of DeepSeek has made people realize that training excellent AI large-language models may not require as much computing power, and the bubble of NVIDIA seems to be bursting.

 

As of the time of writing, NVIDIA has dropped nearly 10% in pre-market trading, and the NASDAQ 100 Index futures have dropped more than 3%.

 

With DeepSeek being so popular, what can we buy in the US stock market? After looking through the materials, I found Kingsoft Cloud, a Chinese-listed stock in the US market. The connections between Kingsoft Cloud and DeepSeek are reflected in the following aspects:

 

Luo Fuli, a key developer of DeepSeek, joined Xiaomi's large-language-model team. Kingsoft Cloud has a close relationship with Xiaomi Group and will continue to provide cloud services to Xiaomi Group in the next three years. Luo Fuli's role in Xiaomi's AI large-language-model project may indirectly increase Xiaomi's demand for Kingsoft Cloud in AI-related cloud services and other aspects.

 

Secondly, Kingsoft Office, a subsidiary of Kingsoft Cloud's parent company Kingsoft Software, has a cooperation with DeepSeek. The intelligent writing function of WPS integrates the DeepSeek-Writer API. To a certain extent, this has created a business intersection between the ecological system where Kingsoft Cloud is located and DeepSeek, helping to enhance the influence of the entire Kingsoft system in the AI field and may also bring more business expansion opportunities to Kingsoft Cloud.

 

And, for companies like NVIDIA, Microsoft, and Meta, although the emergence of DeepSeek has had an impact on the US stock AI field in the short term, in fact, I think DeepSeek is like a catfish thrown into the US stock AI field, which will prompt major giants to update their technologies faster. So, although there is a short-term impact on stock prices, in the long run, it is beneficial.#nvidia $NVDA $ASML $AVGO $MSFT $AMZN $KC