This weekend, both the tech and investment communities were suddenly talking about DeepSeek again. This week, Silicon Valley was on the edge when Chinese artificial intelligence company DeepSeek released its R1 model (DeepSeek R1). In third-party benchmarks, it outperforms leading U.S. AI companies such as OpenAI, Meta, and Anthropic.

In terms of performance, DeepSeek is really good. According to the personal test of the Chinese reporter of the brokerage, the academic literacy and text organization ability of a single essay have been the existence of a group of literary workers. A sense of professional crisis arises spontaneously. From the perspective of its reading and summarizing ability and mathematical logic processing ability, it is also excellent.

The key point is that the cost advantage of this large model is extremely obvious. Therefore, the market also attributed Nvidia's sharp drop on Friday to the popularity of DeepSeek. The market believes that the success of the DeepSeek R1 may have weakened the market's expectations for the demand for NVIDIA's AI chips.

The impact of DeepSeek

China's AI researchers have achieved what many consider to be out of reach: a free and open-source AI model whose performance can rival or even surpass OpenAI's state-of-the-art inference system. What's even more amazing is how they do it: let the AI learn on its own through trial and error, just like humans learn.

In the week before the Spring Festival, China's artificial intelligence company DeepSeek released its R1 model, and Silicon Valley was nervous. In third-party benchmarks, it outperforms leading U.S. AI companies such as OpenAI, Meta, and Anthropic.

The performance figures are impressive. In AIME2024 math benchmarks, DeepSeekR1 has a success rate of 79.8%, surpassing OpenAI's o1 inference model. In standardized coding tests, it demonstrated "expert-level" performance, earning a 2029 Elo rating on Codeforces and outperforming 96.3% of its human competitors.

But what really sets DeepSeek R1 apart is its cost – or rather its low cost. The model costs only $0.14 per million tokens, while OpenAI's cost is $7.50, which is 98% cheaper. Unlike proprietary models, DeepSeek R1's code and training methods are fully open-source under the MIT license, which means that anyone can access, use, and modify the model without restrictions.

Its remarkable reasoning capabilities, which are "pushing the boundaries of open-source AI", caught Silicon Valley off guard and easily became the hottest spot in the tech world this week. The limelight has even overshadowed the Davos Economic Forum, the uncertainty over TikTok and the inauguration of President Trump.

AMD: The new DeepSeek-V3 model has been integrated into the Instinct MI300X GPU

AMD announced that it has integrated the new DeepSeek-V3 model into the Instinct MI300XGPU, which is designed to achieve optimal performance with SGLang. DeepSeek-V3 is optimized for AI inference.

A-share listed companies with equity association, technical cooperation or business synergy with DeepSeek mainly include the following categories, and the specific details of cooperation are subject to the company's announcement and the latest progress:

1. Equity related parties

Zhejiang Dongfang (600120).

Through its Hangzhou Oriental Jiafu Fund, it participated in the DeepSeek angel round investment, and the shareholding path is Zhejiang Oriental →Oriental Jiafu (holding 40% of the shares) → DeepSeek.

The

profit contribution of venture capital business increased to 18% (2023).

Huajin Capital (000532).

The investment platform of Zhuhai State-owned Assets participated in the DeepSeek Pre-A round of financing through Huajin Lingyue Fund to lay out the AI large model track.

2. Computing infrastructure providers

Inspur Information (000977).

Provide AI server clusters for DeepSeek Beijing Yizhuang Intelligent Computing Center, supporting NVIDIA H800 chips and self-developed AIStation management platform.

Sugon (603019).

Contracted the construction of the liquid cooling system of DeepSeek Hangzhou Training Center, with a power density of 35kW and a PUE

Runze Technology (300442).

Provide 3000+ cabinet resources in Langfang data center, and use indirect evaporative cooling technology to reduce operating costs by 15%.

3. Partners in vertical fields

Tors (300229).

Jointly developed a large financial public opinion model with DeepSeek, and has deployed an intelligent research report generation system in CITIC Securities and other institutions.

iFLYTEK (002230).

The education scenario is connected to the DeepSeek-Math model, and the AI math tutoring application "Spark Student" is launched.

Kingsoft Office (688111).

The WPS intelligent writing function integrates the DeepSeek-Writer API, which improves the efficiency of document generation by 3 times and reduces the error rate by 90%.

Fourth, the technical cooperation potential unit

Philips (300287).

The MLA (Multi-Head Latent Attention Mechanism) is used to innovate with the DeepSeek-V2 architecture, and there is potential for technical synergy.

Nanwei Software (603636).

The development of biometric authentication technology based on the multi-head attention mechanism is in line with the direction of DeepSeek technology.

Hangjin Technology (000818).

DeepSeek provides optical modules and switches for DeepSeek, an elite partner of NVIDIA GPUs and networking devices.

5. Data and Ecosystem Partners

Daily interaction (300766).

The second shareholder of High-Flyer Quant (the parent company of DeepSeek), which provides user behavior data to support model training.

Zhuochuang Information (301299).

Cooperate with High-Flyer Quant financial corpus, data resources or for DeepSeek model optimization.

6. Other potential beneficiary directions

AI chips and hardware

Domestic AI chip manufacturers such as Cambrian and Jingjiawei may benefit from the growth in computing power demand driven by DeepSeek technology.

(Redirected from: Little Doctor of Finance).