Track Hyper | DeepSeek: The Potential Logic of Computing Power Demand

Author: Zhou Yuan / Wall Street News

On January 27, NVIDIA's U.S. stock pre-market plummeted by more than 12% (as of January 27, 20:24 Beijing time), falling below the intraday low of $126.86 on December 17, 2024.

Jensen Huang probably never imagined that the force shaking NVIDIA's industrial logic would come from a Chinese private equity firm.

On January 20, the top Chinese private equity firm with a fund size exceeding 100 billion yuan—Huanfang Quantitative—launched the inference large model DeepSeek-R1 (hereinafter referred to as DS-R1) under its subsidiary "DeepSeek," which shocked Silicon Valley with its ultra-low training costs that surpassed industry expectations and performance comparable to OpenAI's o1 official version (in mathematics, code, natural language reasoning, etc.).

The emergence of this GenAI inference large model is essentially a declaration that Chinese technology has reached the forefront of global innovation. From now on, China will transition from a follower to a player capable of setting paradigms in the field of cutting-edge technology.

This marks the true beginning of the GenAI era, with the unveiling being led for the first time by a Chinese technology company.

However, looking to the future, NVIDIA's industrial logic may not be as negatively impacted as its sharply falling stock price suggests; rather, it is likely to be further strengthened.

How should this be understood?

DS-R1 "Instantly Breaks" NVIDIA's Value Chain

On January 24, DS-R1 ranked third in the overall large model category on the model ranking platform Arena.

In the style control model (StyleCtrl) category, DS-R1 tied for first place with OpenAI's o1, achieving a score of 1357, slightly surpassing OpenAI's o1 score of 1352.

OpenAI's o1 is a new model officially released by OpenAI on September 12, 2024, and is the first product in OpenAI's next-generation "inference" model lineup.

DS-R1 achieved performance that is nearly indistinguishable from OpenAI's o1 (some open-source small models even outperform o1), with costs only 3%-5% of the latter (approximately $5.5 million).

This has shattered existing perceptions about the conditions required to achieve cutting-edge AI performance, effectively fragmenting NVIDIA's value chain in an instant.

Currently, OpenAI's o1 has launched the ChatGPT Pro plan, with a monthly fee of up to $200 (approximately 1,440 yuan, totaling an annual fee of 17,280 yuan), which limits ChatGPT Plus or Team accounts to 50 accesses per week.

If using the API for the o1 preview version, the cost is $15 per million tokens for input (approximately 108 yuan) and $60 per million for output (approximately 432 yuan).

Spending about 17,000 yuan allows for only approximately 2,600 inquiries per year?

Such application costs are not affordable for the average person.

Looking at ChatGPT-4o, the personal version has a monthly fee of about $20 (approximately 145 yuan, totaling an annual fee of 1,740 yuan), which also has inquiry limits In contrast, the DS-R1 personal version is free to use; according to official news, the pricing for the DS-R1 API service is RMB 1 per million tokens for input (cache hit) or RMB 4 (cache miss), and RMB 16 per million tokens for output.

Around the world, there is a love for low-cost yet high-performance applications in RMB. Observing such a stark price difference, the American people have made their own choice: currently, the DS-R1 has soared to become the most downloaded popular GenAI model on HuggingFace.

The DS-R1 ranks fourth on the free chart in the Apple App Store in the U.S., surpassing American GenAI applications like Google Gemini and Microsoft Copilot.

In terms of technological iteration, the DS-R1 operates under a fully open-source model, allowing developers and researchers to participate more broadly, harnessing global technological power to promote ecological technological innovation and foster healthy market competition; OpenAI's o1, on the other hand, is a closed-source model that can only rely on the efforts of the OpenAI team, forming a strong support for NVIDIA's industrial logic.

It is foreseeable that companies with closed-source models that do not achieve the performance of the DS-R1 will face mass closures (if they do not want to close, they should join the DS-R1 open-source ecosystem, which is called "if you can't beat them, join them"); even OpenAI's closed-source API value will be restructured.

Currently, the robust performance, low-cost characteristics, and open-source model of the DS-R1 enable commercial organizations and individuals worldwide who aspire to participate in the GenAI ecosystem to significantly promote the breadth and depth of ecological applications.

It can be said that the DS-R1 will become a catalyst for GenAI ecological applications, truly ushering the world into the GenAI application era. However, this time, the market driving force of the new GenAI era has, for the first time, the presence of a Chinese company.

In existing perceptions, while Chinese companies can significantly reduce product costs, they have not been able to lead globally in technology. This time, the DS-R1 launched by DeepSeek breaks the old model of "cheap but lacking in technological content," achieving a new technological era's core driving force for the first time.

Chinese technological strength has shown brilliance in the GenAI technology wave. DeepSeek is not fighting alone.

Previously, the widely popular Doubao large model also launched an iterative version.

According to the Doubao WeChat public account, the Doubao large model version 1.5Pro was officially released on January 22.

This large model leads globally in several public evaluation benchmarks in knowledge, code, reasoning, and Chinese, with a comprehensive score better than top Silicon Valley models like GPT-4o, Claude3.5, and Sonnet.

Has NVIDIA's logic come to an end?

The company that produced the DS-R1, DeepSeek, is a startup and is not publicly listed.

However, on January 27, the listed companies behind the institutions participating in DeepSeek's angel round and Pre-A round financing opened with a "limit-up" increase.

Public information shows that Hangzhou Dongfang Jiafu Fund participated in DeepSeek's angel round investment. Hangzhou Dongfang Jiafu Fund is affiliated with "Zhejiang Dongfang"; the listed company behind Huajin Lingyue Fund, which participated in DeepSeek's Pre-A round financing, is "Huajin Capital." These two companies hit the daily limit up as soon as the A-shares opened on January 27.

However, the differentiation in price movements is evident for Inspur Information and Zhongke Shuguang, which provide foundational computing power support and infrastructure for DeepSeek, as well as companies like Tuolsi that have product technology cooperation with DeepSeek, since the market is still in the process of constructing logic.

To briefly explain the cooperation direction between Inspur Information, Zhongke Shuguang, and DeepSeek: the former provides AI server clusters (NVIDIA H800) to DeepSeek, while the latter is responsible for constructing the liquid cooling system for DeepSeek's training center in Hangzhou.

The A-share companies related to the DS-R1 concept this time are not just a reflection but have direct relevance.

The companies that have a mapping relationship this time are based on NVIDIA's industrial logic.

The essence of DS-R1 achieving strong performance at a low cost lies in improving computing power utilization, rather than eliminating the need for computing power or reducing its demand.

The launch of the DS-R1 with such strong performance by DeepSeek will show the industry that enhancing computing power efficiency is realistic and achievable; after absorbing DeepSeek's ideas and methods for improving computing power efficiency, the industry will scale up computing power to achieve greater benefits.

Therefore, the DS-R1 has not sounded the death knell for NVIDIA's industrial logic; on the contrary, it further solidifies the foundation of "Da Chain" in the secondary market.

This can be summarized with a formula:

DeepSeek's method for improving computing power efficiency + larger scale of computing power = benefits without ceiling limits

The low-cost, high-performance DS-R1 will greatly accelerate the prosperity of the GenAI application ecosystem.

From this perspective, the participation of a vast number of developers and application parties (including B-end commercial organizations and C-end individuals) will further enhance the demand for computing power suppliers.

This is similar to how the introduction of the open-source Android system into the mobile application ecosystem significantly increased the performance demands for mobile SoC chips and storage chips.

In December 2024, when DeepSeek launches the DS-V3, some commentators believe that DeepSeek-V3 will support private deployment and autonomous fine-tuning, providing far greater development space for downstream applications than the closed-source model era. In the next year or two, we are likely to witness a richer array of inference chip products and a more prosperous LLM application ecosystem.

In the history of human development, regardless of the technological era, the ultimate goal has always been one, or rather, only by achieving this goal can the technological era have true driving force.

This goal is benefits.

Commercial value, in essence, is merely a bridge to reach the other side of benefits.

GenAI technology is no exception.

In the past two years in Silicon Valley (taking November 30, 2022, when OpenAI launched ChatGPT-3.5 as a time benchmark), the investment of hundreds of billions of dollars in capital expenditure was not aimed at achieving an "altruistic" ideal, but rather for benefits—win-win does not deny benefits, but rather shares them equally.

Since Steve Jobs launched the first-generation iPhone in 2007, only after truly fostering a prosperous application ecosystem can the enormous benefits that exceed imagination roll in Therefore, the true value of the DS-R1 lies in allowing global GenAI participants to first see and then learn how to enhance computing power utilization efficiency at a lower cost.

Before the launch of the DS-R1 or DS-V3, the performance achieved from spending 1 yuan on computing power resulted in a return of 1.5 yuan. Now, with just 0.1 yuan, one can achieve the same performance as before with an investment of 1 yuan, realizing the same return.

At this point, what will you do?

Will you decide to stop investing, or will you choose to increase your efforts?

How to choose is self-evident.

Therefore, the significant decline in the stock prices of NVIDIA and the A-share "Da Lian" (such as copper cable connections) sector this time, if it is a huge "pit," also shines with the luster of some rare metals