According to the Zhitong Finance APP, Xinda Securities released a research report stating that both algorithm innovation and computing power resources are indispensable on the road to AGI. As early as 2024, the market clearly concluded that to achieve prosperity in the reasoning market, reasoning costs must be reduced. Now, the relevant work of Doubao and the DeepSeek research team has successfully significantly lowered reasoning costs, advancing the industrialization of AI. The success of DeepSeek represents a leap for open-source models compared to closed-source models. From the perspective of the reasoning market, reasoning is expected to scale rapidly, with various functions of text-to-text, text-to-video, and image-to-video likely to iterate quickly, enabling AI to truly understand the physical world, with the singularity of autonomous driving, humanoid robots, and AI sparking industrial revolutions in various industries approaching.
<h2>The main points of Xinda Securities are as follows:</h2>
DeepSeek sets a milestone, exploring the value potential of algorithm innovation
DeepSeek V3 is still a model based on the Transformer architecture, a powerful mixture of experts (MoE) language model with a total of 671 billion parameters, with 37 billion parameters activated for each token. The main contributions of DeepSeek-V3 include: innovative load balancing strategies and training objectives at the architectural level, significantly improved training efficiency at the pre-training level, and knowledge distillation of DeepSeek-R1 at the post-training level. The DeepSeek team completed the pre-training of DeepSeek V3 with an economic cost of only 2.664 million H800 GPU hours on 14.8 trillion tokens. DeepSeek R1 is a culmination based on the architecture of DeepSeek V3, with performance benchmarked against OpenAI-o1. DeepSeek-R1-Zero is a model trained through large-scale reinforcement learning (RL) without supervised fine-tuning (SFT) as a preliminary step, demonstrating exceptional reasoning capabilities.
Through reinforcement learning, DeepSeek-R1-Zero naturally exhibited many powerful and interesting reasoning behaviors. Among them, DeepSeek-R1-Zero demonstrated self-verification, reflection, and the generation of long chains of thought (CoT), marking an important milestone in the research community. Notably, this is the first validated open research that can purely incentivize the reasoning capabilities of LLMs through RL without SFT. This breakthrough is expected to pave the way for future developments in the field. Additionally, the DeepSeek team has open-sourced DeepSeek-R1-Zero, DeepSeek-R1, and models based on (1.5B, 7B, 8B, 14B, 32B, 70B), which outperform OpenAI o1-mini in some tests.
OpenAI sounds the counterattack horn, releasing o3-mini and Deep Research in succession
In response to the challenge from DeepSeek, OpenAI released O3-mini, significantly enhancing its ability to understand the physical world and programming capabilities. O3-mini demonstrated astonishing strength in high-difficulty challenges of physical simulation, exhibiting stronger physical reasoning capabilities when simulating the rotation of a small ball, while DeepSeek R1 showed anti-gravity phenomena
In addition, the o3-mini can generate a bouncing program for small spheres in four-dimensional space, demonstrating high potential.
Furthermore, OpenAI has released Deep Research, a new agent from OpenAI—just one prompt, and ChatGPT will search, analyze, and synthesize hundreds of online resources to create comprehensive reports at the level of research analysts. It is powered by the upcoming version of the OpenAI o3 model, which is optimized for web browsing and data analysis, utilizing reasoning to search, interpret, and analyze vast amounts of text, images, and PDFs on the internet, adjusting as needed based on the information encountered.
The future of AI still has vast oceans and stars to explore, the commercialization singularity of AGI is approaching.
The Google DeepMind team has classified AI into six major development stages. From a narrow perspective, there are already AI models in specialized fields that can achieve superhuman levels. Models like AlphaFold, AlphaZero, and StockFish can completely surpass humans in specific domains. However, from the perspective of general artificial intelligence, AI development is still at a relatively low level, with ChatGPT only classified as Level 1-Emerging. FIGURE is one of the leaders in the humanoid robot field, and from its perspective, one can glimpse the historical opportunities of AI. Currently, 50% of the global GDP's labor force consists of human labor, representing a space of approximately $42 trillion; just this aspect alone presents a vast prospect for AI.
Investment Advice: Algorithm innovation and computing power investment complement each other; it is recommended to focus on AI industry chain targets.
The development speed in the AI field surpasses that of traditional manufacturing. Since 2023, Scaling Laws have begun to take effect, initiating a &#34;arms race&#34; for computing resources globally. Motivated by factors such as geopolitical issues, the pursuit of algorithms has finally been heralded by Chinese teams, which have relatively scarce computing power. On the road to AGI, both algorithm innovation and computing resources are indispensable. As early as 2024, it was concluded that to achieve prosperity in the reasoning market, reasoning costs must be reduced. Now, the relevant work of Doubao and the DeepSeek research team has successfully significantly lowered reasoning costs, advancing the industrialization of AI. The success of DeepSeek represents a leap for open-source models compared to closed-source models.
Indeed, within a very limited timeframe, North American tech giants may focus their limited energy on the algorithm level to fully exploit their computing resource potential. However, this does not mean that investment in computing power will cease. On the contrary, the industrialization of AI is expected to accelerate computing power investment in the medium to long term, avoiding the pitfalls of previous AI booms that ultimately failed due to inability to land.
From the perspective of the reasoning market: Reasoning is expected to scale rapidly, with various functions such as text-to-text, text-to-video, and image-to-video likely to iterate quickly, enabling AI to truly understand the physical world. The singularity of autonomous driving, humanoid robots, and AI sparking industrial revolutions across various industries is approaching. From the perspective of the training market: On one hand, cutting-edge exploration of training models still requires substantial computing power investment, and the scaling of the reasoning market is expected to give rise to new directions for model exploration
On the other hand, research progress in areas such as world models is expected to accelerate. However, it is also observed that the elimination of large model vendors will accelerate, with closed-source large model vendors unable to surpass open-source models or quickly clear out.
Zuckerberg stated at a Meta earnings call, &#34;Over time, just as every business has a website, a social presence, and an email address, in the future, every business will also have an AI agent that customers can interact with. Our goal is to integrate this into an AI agent to drive sales and save costs.&#34; This moment is gradually approaching. Historically, when computers moved from laboratories to households, it did not lead to the decline of related companies but rather gave birth to many great enterprises. Currently, AI is also expected to undergo this process, and it is recommended to pay attention to relevant targets in the industrial chain.
Related Stocks
AI Cloud Side: Industrial Fulian (601138.SH), Hu Dian Co., Ltd. (002463.SZ), Shengyi Technology (600183.SH), Shenzhen South Circuit (002916.SZ), Shenghong Technology (300476.SZ), Cambricon (688256.SH), Haiguang Information (688041.SH).
AI End Side: Lens Technology (300433.SZ), Lingyi Technology (002600.SZ), Pegatron (002938.SZ), Dongshan Precision (002384.SZ), Espressif Technology (688018.SH), Rockchip (603893.SH), Hengxuan Technology (688608.SH), Allwinner Technology (300458.SZ), Zhaoyi Innovation (603986.SH), Amlogic (688099.SH), etc.
Risk Factors
Macroeconomic downturn risk; downstream demand falling short of expectations risk; intensified Sino-U.S. trade friction risk

DPSK

Cinda Securities released a research report stating that DeepSeek-R1 is an important milestone towards AGI, with algorithm innovation and computing power resources being key to achieving AGI. The DeepSeek team is promoting the industrialization of AI by reducing inference costs, and the DeepSeek V3 model has shown significant improvements in training efficiency. DeepSeek-R1-Zero demonstrates exceptional reasoning capabilities through reinforcement learning, indicating that an industrial revolution in fields such as autonomous driving and humanoid robots is imminent

Zhitong

DeepSeek

Lens

ALLWINNERTECH

YIWU CCC

Dynanonic

Rockchip

Destiny Tech100

Alphabet

Avary Holding

Sunwave

Meta Platforms

SENSETIME-W

Suochen

Agnico Eagle Mines

iShares Self-Driving EV&Tech ETF

Huazhi Digital Media

Amlogic

SIASUN

Tongxing

Cinda Securities

Skychem Technology

OpenAI

GIGADEVICE

LY iTECH

PIA Automation

DSBJ

Global X Autonomous & Electric V

GX CN SEMICON

IFLYTEK

UBTECH ROBOTICS

Cambricon

Hbmc

Sytech

Alphabet - C

Espressif Systems

Cinda Securities: DeepSeek-R1 is a milestone for AGI, which is a long-term positive for computing hardware