Guosheng Securities: AI Capability Boundaries Will Usher in a New Round of Expansion After the Festival, the main focus is on two directions

The Zhitong Finance App learned that Guosheng Securities released a research report saying that as the RL algorithm gradually replaces the autoregressive algorithm in the pre-training stage and uses more powerful computing power and more data to generate a thought chain based on the RL algorithm training model, they will jointly form a new AI scaling law. Algorithmic innovation and computing power will run wild together on this curve, and AI's ability boundaries will usher in a new round of expansion. During the Spring Festival holiday, in addition to DeepSeek, incidents such as the implementation of North American tariffs also occurred. Combined with the market's perception of marginal changes in DeepSeek, it was more responsive to rapid deployment on the application side, so the main offensive direction after the holiday season was mainly focused on the end side and domestic alternative chains.

Looking at the medium to long term, global computing power supporting industry chains such as communications and energy will also usher in a new period of development. After the holiday, it is recommended to focus on: end side: IoT module leader, Meige Intelligence (002881.SZ), Yiyuan Communications (603236.SH), Guanghetong (300638.SZ); domestic alternatives: Cambrian-U (688256.SH), SMIC (688981.SH,00981), ZTE Communications (000063.SZ).

Guosheng Securities's main views are as follows:

What does DeepSeek do

The new scaling law has been verified: Over the past few years, the growth of AI models has mainly relied on large-scale accumulation during the pre-training stage, increasing the scale and capabilities of models through the continuous accumulation of existing human data. However, as training exhausted existing human data, the scaling law in the pre-training phase gradually slowed down due to illusions and diminishing marginal benefits caused by the path of superimposing synthetic data.

Since 2024, the new scaling law based on reinforcement learning has become a key direction in the industry, and excellent models such as GPT-O1 and Deepseek R1 have emerged successively. RL's new AI growth curve initially showed a very high investment cost ratio. This is the fundamental reason why Deepseek V3 rapidly evolved into R1 and showed strong capabilities. Currently, RL is still based on traditional pre-training models. Adding RL during the inference stage makes the model more powerful in reasoning skills when facing science problems.

Looking forward to the future, as RL algorithms gradually replace autoregressive algorithms in the pre-training stage, and use more powerful computing power and more data to generate thought chains based on the RL algorithm training model, they will jointly form a new AI scaling law. Algorithmic innovation and computational power will run wild together on this curve, and AI's ability boundaries will usher in a new round of expansion.

Extreme engineering optimization: DeepSeek's true innovation lies in extreme engineering optimization, which mainly relies on such things as key value caching, innovative MoE architecture, and extreme compression of Nvidia's GPU efficiency based on PTX layer assembly language. Through these innovative engineering optimizations, DeepSeek has broken the inherent cost reduction cycle curve of Western model companies, and can use a lower price to provide a user experience close to the head model and improve the efficiency of using global computing power.

Generous open source: Unlike the Western model business concept of OAI and Anthropic, which is gradually moving towards closed source, DeepSeek has open sourced its own innovative principles and models. On the one hand, it enables global model makers to share the cost performance improvements brought about by new engineering methods, and on the other hand, it enables global users to deploy locally or through public clouds to avoid high premiums. This is the core reason why Deepseek has received such high praise from global developers.

DeepSeek's impact on Nvidia

The advent of Deepseek has brought the world one step closer to the realization of AGI. Guosheng Securities believes that the overseas layout in the field of computing power will not slow down because of Deepseek; on the contrary, because of the advent of Deepseek, it will further strengthen the momentum for global tech giants, and overseas tech giants will further increase their layout in the field of computing power. Specifically, on the one hand, it may further increase Nvidia's GPU procurement efforts, and on the other hand, it will also step up the progress of self-developed ASIC solutions. Furthermore, the US government may further tighten chip export restrictions and try to carry out a final blockade at the computing power level to limit AI development in other countries and regions and maintain its so-called AI leadership position.

As far as Nvidia is concerned, Guosheng believes that Deepseek's phased victory will continue to drive overall demand in the computing power market. It is not believed that Nvidia's demand and orders will have any significant impact. On the other hand, Nvidia may further exceed expectations. On the other hand, Nvidia may also accelerate the iteration speed of its next-generation products, and performance is expected to achieve major breakthroughs in extremely innovative technology routes (such as CPO, OIO, etc.), which are expected to continue to accelerate. The fall in Nvidia's stock price is not a reflection of changes in demand, but more of an uncertainty caused by concerns about possible further tightening of chip export restrictions.

DeepSeek's impact on Chinese computing power investment

Regarding domestic computing power: After the launch of DeepSeek, the US was under competitive pressure, and calls for China to further strengthen computing power sanctions became stronger. At the same time, due to the open source and low cost characteristics of DeepSeek, the cost performance ratio and ROI of domestic video cards used for inference have risen sharply. At this point, based on the combination of SMIC's manufacturing capabilities, the design capabilities of chip manufacturers, and the application ability of the Deepseek model, China's flywheel with autonomous and controllable computing power has begun to spin.

For the edge side: DeepSeek has two main points of traction on the edge side. The first is the reduction in the price of advanced models in the cloud, which reduces the use and development and operation costs of AI applications and AI hardware, which will promote the expansion of AI hardware. The second is DeepSeek's ability to optimize small models. This time, Deepseek also released small models after some optimization based on Qwen and Llama. Compared with previous small models, the ability has improved to a certain extent. As RL's new growth curve continues to expand, and edge computing power increases, edge reasoning will accelerate into reality, and edge computing power is expected to accelerate.

Deepseek's impact on global computing power investment

GPU: DeepSeek's impact on GPUs is positive, because at present, whether from a training or reasoning perspective, the advantages of GPUs are very obvious. Computing power will continue to grow exponentially in the long term. We are still optimistic about Nvidia's definitive advantage in the field of AI infrastructure, and at the same time, we are optimistic about its acceleration in new technology fields that can greatly improve performance, and the accelerated implementation of CPO and OIO. It is recommended to focus on: Tianfu Communications, Corning, Taichen Guang, etc.

ASIC: The positive impact of reinforcement learning on reasoning may accelerate the demand and progress of large model teams and cloud vendors for inference and self-developed computing power. The final market depends on the comparison of unit computing power costs. Currently, Nvidia still has a clear advantage in the field of inference, but cloud vendors may continue to spend more to firmly promote self-developed computing power. ASIC partners that benefit cloud vendors and their supporting industry chains: Broadcom, AMD, MRVL, ARISTA, CLS, Zhongji Xuchuang, Xinyisheng, Cohr, Lite, etc.

Communications: Deepseek's impact on communication is quite comprehensive. Generally, the communication requirements of training clusters are higher than inference clusters of the same size, but considering cost performance issues, many training clusters now also take into account other functions such as recommendation algorithms, reasoning, scientific research, financial transactions, etc., and training clusters will still steadily develop in the direction of larger scale. At the same time, inference requirements need to consider scale effects. It is expected that the inference market dominated by cloud vendors will still dominate. Currently, the computing power on the end side is limited, which will not have much impact. It is expected that inference clusters will continue the development trend of the past.

Overall, whether it is training or reasoning, more high-end switches and optical modules will still be used. We tend to recommend targets with superior products, sufficient orders in hand, and smooth capacity expansion. It is recommended to focus on: Broadcom, Arista, CLS, Zhongji Xuchuang, Xinyisheng, Cohr, Lite, Ciena, and Decolly.

Energy: Overseas energy-related support is poor. It is expected that the emergence of Deepseek will strengthen the determination of overseas technology giants to accelerate the deployment of infrastructure. The energy sector continues to be optimistic about the overall direction. It is recommended to focus on SMR, OKLO, Weiteng Electric, and electrical alloys.

Risk warning: AI development falls short of expectations, demand for computing power falls short of expectations, market competition risks.