DeepSeek drives affordable high-quality models with its combination of low cost, high performance, and strong reasoning

1. DeepSeek models are updated intensively, and the number of users will continue to grow rapidly
2. Since 2024, DeepSeek has rapidly risen and iterated in the field of AI. From the end of December 2024 to the end of January 2025, updates were particularly intensive, releasing V3 with many parameters and improved performance, R1 that supports thought chain output and model training, and visual and multimodal models that focus on the image field. From the end of December 2024 to the end of January 2025, the number of global users surged from 347,000 to 119 million. Compared with ChatGPT, DeepSeek reached the user scale of ChatGPT in just over a year, and topped the list of average monthly active users in China in January, with a significant increase in APP downloads.
3. DeepSeek has three characteristics: low cost, high performance, and strong reasoning
4. DeepSeek drives model parity, and recommends focusing on investment opportunities in computing power, AI applications, and end-side

1. Computing power: With more users using DeepSeek and more AI applications emerging in the future, the demand for computing power is growing exponentially. Although the model efficiency has been improved with the advancement of AI technology, the growing number of users and applications has put forward higher requirements for computing resources, and consumption has also increased dramatically. 
2. B-side applications: AI Agent is comprehensively reconstructing traditional SaaS applications. Compared with the traditional knowledge base structured management model, AI Agent's vector database has a strong self-learning ability, can automatically understand the content of documents, achieve more efficient knowledge management, and provide strong support for the digital transformation of enterprises. C-side applications: As an important commercial application of generative AI, AI Agent has been widely used in many industries such as e-commerce, education, tourism, hotels, and customer service. 
3. End-side: AI is influencing the world in terms of content, applications, hardware, and ecology. AI Agent has moved from "digital" to "embodied"; with the development of the market, large models are more widely connected to hardware products, and the coordinated development of software and hardware is the key to future competition.
4. 2. Low cost: DeepSeek is in the best price/performance range of models , which is significantly lower than similar models such as OpenAI
5. 　　DeepSeek-V3 has greatly improved model efficiency through algorithm innovation and engineering optimization, thereby reducing costs and improving cost-effectiveness. 1) From the perspective of algorithm innovation, DeepSeek-V3 adopts the self-developed MoE architecture, with a total parameter volume of 671B, and each token activates 37B parameters, achieving multi-dimensional benchmarking against GPT-4o. Its sparse expert model MoE has been expanded to 256 routing experts plus 1 shared expert. Each token activates 8 routing experts and is sent to up to 4 nodes. It also introduces a redundant expert deployment strategy to achieve load balancing between different experts in MoE during the reasoning phase, and also proposes a load balancing strategy without auxiliary loss to reduce performance degradation. In addition, the multi-head attention mechanism MLA revolves around the video memory, bandwidth and computing efficiency of the reasoning phase. By innovating the underlying software architecture, introducing mathematical transformations to reduce the memory usage of the kv cache, alleviate the video memory and bandwidth bottlenecks during transformer reasoning, and optimize the attention calculation method to further improve efficiency. At the same time, the innovative training target MTP is adopted to allow the model to predict multiple future tokens at one time during training, expand the prediction range, enhance the ability to understand the context, optimize the training signal density, and increase the reasoning speed by 1.8 times. 2) In terms of engineering optimization, DeepSeek-V3 innovatively implements FP8 + mixed precision strategy on a large scale, reducing the calculation accuracy from the mainstream FP16 to FP8, retaining the mixed precision strategy, and retaining FP16/32 in important operator modules to ensure accuracy and convergence, taking into account model stability and reducing computing power costs. 3) In solving the communication bottleneck problem, the DualPipe efficient pipeline parallel algorithm is adopted to achieve communication overhead close to 0.
6. 　　A series of innovations and optimizations have made the training cost of DeepSeek-V3 only $5.57 million, and it took less than two months. According to the paper, the formal training cost of DeepSeek-V3 is only $5.576 million. In the pre-training stage, it only takes 180,000 H800 GPU hours to train one trillion labeled DeepSeek-V3, which is only 3.7 days on the 2048 H800 GPU cluster owned by DeepSeek. Adding 2.664 million GPU hours of pre-training, 1.19 million GPU hours of context length expansion, and 5,000 GPU hours of post-training, it is concluded that the complete training of DeepSeek-V3 only requires 2.788 million GPU hours. Assuming the rental price of H800GPU is $2 per GPU hour, the total training cost is only $5.576 million.
7. 　The cost of DeepSeek's general and inference models has dropped significantly compared to similar models such as OpenAI . 1) General model : The DeepSeek-V3 model API service pricing is adjusted to 0.5 yuan (cache hit) / 2 yuan (cache miss) per million input tokens, and 8 yuan per million output tokens. In addition, the V3 model has a 45-day preferential price experience period: Before February 8, 2025, the API service price of V3 will remain at 0.1 yuan (cache hit) / 1 yuan (cache miss) per million input tokens, and 2 yuan per million output tokens. 2) Inference model : The DeepSeek-R1 model API service pricing is 1 yuan (cache hit) / 4 yuan (cache miss) per million input tokens, and 16 yuan per million output tokens.
8. 　　Table 2 : Pricing comparison of different large model API services

9. Large model name	API Service Pricing
   DeepSeek	DeepSeek-V3: 0.5 yuan (cache hit) / 2 yuan (cache miss) per million input tokens, 8 yuan per million output tokens DeepSeek-R1: 1 yuan (cache hit) / 4 yuan (cache miss) per million input tokens, 16 yuan per million output tokens (halved during the promotion period before February 9)
   ChatGPT	GPT-4 Turbo: Input price is about 70 yuan per million tokens o3-mini: Price is 63% lower than the previous generation
   Thousand Questions on Tongyi	Text model: qwen2-72b-instruct: input price is 0.005 yuan/1,000 tokens, output price is 0.01 yuan/1,000 tokens qwen1.5-110b-chat: input price is 0.007 yuan/1,000 tokens, output price is 0.014 yuan/1,000 tokens qwen-72b-chat: input and output prices are both 0.02 yuan/1,000 tokens Visual understanding model: Qwen-VL-Plus: input price is 0.0015 yuan/1,000 tokens Qwen-VL-Max: input price is 0.003 yuan/1,000 tokens
   A Word from the Heart	It will be completely free from 0:00 on April 1st
   Bean buns	Post-payment mode: Taking Doubao general model pro-32k as an example, the inference input is 0.0008 yuan/thousand tokens, the inference output is 0.002 yuan/thousand tokens, and the comprehensive price of model inference is 0.001 yuan/thousand tokens. Pre-payment mode: Taking Doubao general model pro-32k as an example, the monthly price of 10K TPM is 2,000 yuan, and the average price is 0.0046 yuan/thousand tokens.
   like	Open platform multimodal image understanding model: moonshot-v1-8k-vision-preview: 12 yuan per 1M tokens moonshot-v1-32k-vision-preview: 24 yuan per 1M tokens moonshot-v1-128k-vision-preview: 60 yuan per 1M tokens Context cache: Cache creation fee: 24 yuan/M token Cache storage fee: 5 yuan/M token/minute Cache call fee: 0.02 yuan/time

10. 　　Source: Yuanda Information Securities Research Institute
11. 　　3. High performance & strong reasoning : Deepseek has outstanding algorithm capabilities and its model performance ranks among the best in the world
12. 　　DeepSeek-R1 inherits the innovative architecture of V3 and uses reinforcement learning technology on a large scale in the post-training stage. It automatically selects valuable data for annotation and training, reduces the amount of data annotation and waste of computing resources, and greatly improves the model reasoning ability when there is only a small amount of annotated data. In tasks such as mathematics, code, and natural language reasoning, DeepSeek achieved a pass@1 score of 79.8% in the AIME 2024 evaluation, slightly exceeding OpenAI-o1; on MATH-500, it scored 97.3%, which is comparable to OpenAI-o1 and significantly better than other models .
13. 　　DeepSeek's distillation technology significantly improves the reasoning ability of small models . According to the technical documentation of DeepSeek-V3, the model uses high-quality data generated by data distillation technology to improve training efficiency. A small amount of high-quality data is synthesized through existing high-quality models as training data for new models, thereby achieving an effect close to training on the original data. DeepSeek has released distilled versions of R1 ranging from 1.5 billion to 70 billion parameters. These models are based on architectures such as Qwen and Llama, showing that complex reasoning capabilities can be encapsulated in smaller and more efficient models. The distillation process involves fine-tuning these smaller models using synthetic reasoning data generated by the full DeepSeek-R1, thereby maintaining high performance while reducing computational costs. Let larger models learn high-level reasoning patterns first, and then transplant these results to smaller module packages .
14. 　2. DeepSeek drives the model to become more affordable. We recommend paying attention to investment opportunities in computing power, AIAgent , and the end-side.
15. 　　1. DeepSeek drives the model to become more affordable , and computing power requirements increase significantly
16. 　　The popularity of DeepSeek has brought attention to the economic term "Jevons Paradox". The "Jevons Paradox" was proposed by economist William Stanley Jevons in 1865. At that time, Britain was worried about the possible depletion of coal resources, and people believed that improving the efficiency of coal use could alleviate resource shortages. However, Jevons pointed out that the efficiency improvement brought about by technological progress will lead to increased resource consumption. For example, when the efficiency of coal power technology is improved, it means that more energy can be obtained at a lower cost, which will promote the rise of more industries that rely on coal energy, such as factories, trains, ships, etc., which will in turn stimulate a substantial increase in coal demand and accelerate the consumption of coal resources. 
17. 　　The popularity of DeepSeek has led to more users starting to use AI services, which is like opening the Pandora's box of AI application demand. With more users using DeepSeek and more AI applications emerging in the future, the demand for computing power is growing exponentially. The advancement of AI technology is like the improvement of coal-fired power technology in the past. Although the model efficiency has improved, the growing number of users and applications has put forward higher requirements for computing resources, and consumption has also increased dramatically.
18. 　　The number of DeepSeek users has increased significantly, and it is facing the problem of insufficient server resources. Currently, the website has suspended API recharge, showing "Current server resources are tight. To avoid affecting your business, we have suspended API service recharge. The existing recharge amount can continue to be used, thank you for your understanding." When using it, we also found that the webpage deepseek Q&A often feedbacks "Server busy, please try again later."
19. 　It is recommended to focus on the computing power link centered on domestic computing power and AI reasoning needs, especially computing power supporting industries such as IDC, servers, and domestic chips. Haiguang Information and Inspur Information are recommended.
20. 　　2. AIAgent has a broad market space, and B-end and C-end applications have great potential
21. 　　DeepSeek was quickly integrated into the platforms of various cloud vendors, directly raising the lower limit of model capabilities and accelerating the upgrade of AI application development. At a time when the AI ​​field is booming, DeepSeek-R1, a large model launched by DeepSeek on January 20, has quickly become the darling of major cloud vendors at home and abroad with its outstanding performance in mathematics, code, natural language reasoning and other tasks comparable to the official version of the OpenAI o1 model, as well as the advantages of supporting free commercial use, arbitrary modification and derivative development under the MIT license agreement. As of February 5, domestic mainstream cloud platforms such as Huawei Cloud, Tencent Cloud, Alibaba Cloud, Baidu Smart Cloud, and international cloud giants such as Amazon AWS and Microsoft Azure have announced that DeepSeek will be quickly integrated into their respective platforms. For example, Tencent Cloud deploys the R1 large model to the high-performance application service HAI with one click, and developers can access and call it in just 3 minutes. It also launched a "Developer Gift Pack" to achieve one-click deployment of the entire DeepSeek model series; Alibaba Cloud PAI Model Gallery supports one-click deployment of DeepSeek-V3 and DeepSeek-R1 on the cloud. This extensive and rapid integration directly raises the lower limit of model capabilities. Some complex functions that were previously difficult to implement due to insufficient basic model capabilities can now be easily achieved after integrating DeepSeek. With the help of these cloud platforms integrated with DeepSeek, many companies have greatly shortened the cycle from model building to function implementation when developing AI applications, significantly improved development efficiency, and accelerated the upgrade of AI application development.
22. 　　Table 3 : Domestic and foreign cloud vendors accessing DeepSeek models

23. Manufacturer Name	Access time	Related information
    HUAWEI CLOUD	2025/2/1	Joint Silicon Mobility launched the DeepSeek R1/V3 inference service based on Huawei Cloud Ascend cloud service
    Tencent Cloud	2025/2/2	Supports one-click deployment of DeepSeek-R1 models. Developers can start and configure the model in just 3 minutes.
    Alibaba Cloud	2025/2/3	PAI Model Gallery supports one-click deployment of DeepSeek-V3 and DeepSeek-R1 models on the cloud
    Baidu Smart Cloud	2025/2/3	Qianfan platform officially launched DeepSeek-R1 and DeepSeek-V3 models, and launched ultra-low price plans and limited-time free services
    Volcano Engine	2025/2/4	Full support for DeepSeek series large models, including models of different sizes such as V3 and R1
    JD Cloud	2025/2/4	Officially launched DeepSeek-R1 and DeepSeek-V3 models, supporting public cloud online deployment and dedicated hybrid private instance deployment modes
    Tianyi Cloud	2025/2/5	The DeepSeek-R1 model is fully integrated into its intelligent computing product system
    Microsoft Azure	2025/1/30	Users can deploy DeepSeek-R1 models on Azure AI Foundry and GitHub
    Amazon AWS	2025/1/30	Users can deploy DeepSeek-R1 models in Amazon Bedrock and Amazon SageMaker AI
    Nvidia	2025/1/31	DeepSeek-R1 model is now available on NVIDIA NIM. On a single NVIDIA HGX H200 system, the full version of DeepSeek-R1 671B can process up to 3872 tokens per second.

24. 　　Source: Yuanda Information Securities Research Institute
25. 　　The AIAgent market has a broad space, and there is much room for development on both the B-end and the C-end. According to TouBao Research Institute, the scale of China's AIAgent market reached 55.4 billion yuan in 2023 , and it is expected that by 2028, this figure will climb to 852 billion yuan, with an average annual compound growth rate of 72.7%. Among them, AIAgent in vertical fields has emerged as a new force and quickly became the focus of the technology industry. Industry insiders predict that the scale of the AI ​​agent market in vertical fields is expected to reach ten times that of SaaS, and may even give birth to unicorn companies with a valuation of more than 300 billion US dollars.
26. 　　From the perspective of market application, the value of AIAgent is mainly reflected in two aspects: ToC and ToB . 1) B-side scenario: AIAgent is comprehensively reconstructing traditional SaaS applications. Compared with the traditional knowledge base structured management model, AIAgent's vector database has strong autonomous learning capabilities, can automatically understand document content, achieve more efficient knowledge management, and provide strong support for the digital transformation of enterprises. 2) C-side scenario: As an important commercial application of generative AI, AIAgent has been widely used in e-commerce, education, tourism, hotels, customer service and other industries. Through intelligent interactive services, AIAgent not only improves user experience, but also promotes the upgrading and transformation of traditional industries, bringing more convenient and personalized services to consumers.
27. 　　DeepSeek was quickly integrated into the platforms of various cloud vendors, directly raising the lower limit of model capabilities and accelerating the upgrade of AI application development. Recommended : B-side : Dingjie Digital Intelligence, UFIDA Network; C-side : Kingsoft Office.
28. 　　3. DeepSeek leads the open source technology ecosystem, and its low cost and high performance are conducive to the explosion of edge AI
29. 　　AI is influencing the world in terms of content, applications, hardware, and ecology. AI agents have moved from "digital" to "embodied" . As the market develops, large models are more widely connected to hardware products. The coordinated development of software and hardware is the key to future competition.
30. 
    
31. A.I.P.C.
32. 　　AI PC is expected to become the first stop for the implementation of the end-side model with its high-performance hardware and productivity attributes . The key feature of AIPC is that it can meet the personalized needs of users in a more customized, efficient and secure way by running large models locally. In the early days, AIPC was able to realize functions such as text-to-text, text-to-picture, automatic report generation, and AI local knowledge base, but given the limitations of the model capabilities at the time, a cloud + end hybrid model solution was generally adopted. With the improvement of DeepSeek model performance, its application scope is also constantly expanding. Lenovo, Huawei and other well-known brand manufacturers have keenly captured this technical advantage and have connected Deepseek to their own systems. This move enables users to get a smarter and more convenient AI interaction experience when using devices of these brands. Whether it is document processing in daily office or creative inspiration during leisure and entertainment, users can feel the efficiency and convenience brought by AIPC, enjoy a smoother and more natural human-computer interaction, and truly experience the changes that technology has brought to life.
33. 
    
34. AI Glasses
35. 　　Glasses have become an excellent carrier for the implementation of edge-side AI, which fits a wide range of application scenarios in modern life . Glasses are the objects closest to the three major senses of mouth, ears and eyes among human wearable devices. Technological advances have made them an excellent carrier for the implementation of edge-side AI: AI glasses integrate multiple functions of components such as cameras, glasses, microphones and Bluetooth headsets, and can directly and naturally realize multi-modal input and output of sound, language and vision, which perfectly meets the use conditions of complex AI functions.
36. 　　　Technology companies are making plans one after another, involving more than 50 products. The success of Ray-Ban Meta fully proves the feasibility of AI glasses as an innovative product. Technology companies are making plans one after another, stimulating the vigorous development of the entire industry. According to VR Tuo Gyro, there are currently more than 40 domestic and foreign manufacturers entering the AI ​​glasses market, including Internet giants, mobile phone giants, and AR star companies, and the number of products involved is expected to exceed 50.
37. 　　Table 4 : Technology companies’ AI eye layout

38. Manufacturer Name	Product Name		Release time	Price	Selling Points
    Huawei		Huawei Vision Glass	2024	From 1349 yuan	Lightweight and comfortable, equivalent to a 120-inch giant screen at 4 meters in front, can present a giant screen of over 200 inches in 3D mode, and supports a variety of 3D sources
    Huawei		Huawei Smart Glasses 2	2024	From 1399 yuan	Equipped with Hongmeng system, supporting a variety of intelligent interactive functions
    Huawei		Huawei X GENTLE MONSTER Eyewear II LANG-01	2024	From 889 yuan	Combining stylish design with smart functions
    Millet		Xiaomi AI Glasses	Expected to be in April 2025	Not yet announced	Equipped with AI functions, audio headset module, and camera module, it is fully comparable to Ray-Ban Meta
    Rocks		Rokid Glasses	November 18, 2024	2499 Yuan	Equipped with Qualcomm Snapdragon AR1 platform, 12-megapixel camera, supports high-definition photography and video
    Thunderbird Innovation		Thunderbird V3	January 7, 2025	1799 CNY	Equipped with Falcon imaging system, Tongyi's exclusive customized large model, and the first generation Snapdragon® AR1 flagship chip
    Thunderbird Innovation		Thunderbird X3 Pro	Expected to be in Q2 2025	Not yet announced	Equipped with Firefly light engine, RayNeo waveguide, full-color MicroLED light engine
    Meta		Meta-Ban Ray	September 2023	From $299	Built-in directional speakers, microphones, cameras and other components, can be used for FPV shooting/video recording, calls, listening to music, etc.
    Samsung		Samsung AI Smart Glasses	Expected September 2025	Not yet announced	Equipped with AR1, supports Gemini model
    apple		Apple AI Glasses	Expected 2025	Not yet announced	Reportedly under development, may have AR capabilities
    Lee Weike Technology		Li Wei Ke Meta Lens Chat	Expected to be Q4 2024 or Q1 2025	Not yet announced	We have reached a strategic cooperation with Dr. Glasses and entered Dr. Glasses' offline stores nationwide
    Honeycomb Technology		Jiehuan AI Audio Glasses	September 2024	600-800 Yuan	Provides 8 frames and 14 colors, quick-release structure design, weighs about 30 grams, and has a battery life of 11 hours
    Shanji Technology		Shanji AI "Snap Mirror"	December 19, 2024	From 999 yuan	The first mass-produced AI glasses in China, equipped with Loomo OS, the world's first AI memory system developed by Shanji
    XREAL		XREAL One	2024	From 3299 yuan	Smart AR glasses with high-definition display and intelligent interaction functions
    INSIDE		INMO GO 2	November 29, 2024	From 3999 yuan	All-in-one AI+AR smart glasses with real-time simultaneous translation, portable prompting and other functions
    See Technology		Looktech AI Glasses	2024	Not yet announced	With AI photography, recognition and other functions
    Star Technology		Xingchen AI Glasses	2024	Not yet announced	Equipped with StarCent Technology's AI technology, it has intelligent interactive functions

39. 　　Source: Yuanda Information Securities Research Institute
40. 　　3. Investment advice
41. 　　1) It is recommended to focus on the computing power link centered on domestic computing power and AI reasoning needs, especially computing power supporting industries such as IDC, servers, and domestic chips. Haiguang Information and Inspur Information are recommended.
42. 　　2) DeepSeek is quickly integrated into the platforms of various cloud vendors, directly raising the lower limit of model capabilities and accelerating the upgrade of AI application development. Recommended: B-side: Dingjie Digital Intelligence, UFIDA Network; C-side: Kingsoft Office.
43. 　　3) The improvement of small model capabilities has promoted the deployment of terminal models. We are optimistic about the potential for AI terminals to explode as a new generation of computing platforms. Recommended stocks to watch: iFlytek, Luxshare Precision, and Goertek.
44. 　　Table 5 : Wind consensus earnings forecasts for related companies

45. company	Code	PB	Net profit attributable to parent company (100 million yuan)			ON			Total market value (100 million yuan)
    			2024E	2025E	2026E	2024E	2025E	2026E
    Haiguang Information	688041.SH	15.9	19.5	28.7	39.0	162.3	110.2	81.1	3159
    Inspur Information	000977.SZ	4.9	23.0	28.7	34.3	40.5	32.5	27.2	931
    Dingjie Digital Intelligence	300378.SZ	5.6	1.8	2.2	2.7	66.1	53.5	43.1	118
    UFIDA Network	600588.SH	7.2	-0.4	3.2	6.2	-1553.2	194.3	100.4	625
    Kingsoft Office	688111.SH	16.6	15.2	19.1	24.2	116.5	93.0	73.5	1776
    iFLYTEK	002230.SZ	7.7	6.0	9.6	13.3	211.0	130.9	94.7	1262
    Luxshare Precision	002475.SZ	4.9	135.9	171.8	209.1	23.2	18.3	15.1	3151
    Goertek	002241.SZ	3.0	27.3	36.4	45.3	35.6	26.7	21.5	972

46. 　　Source: Wind, Yuanda Information Securities Research Institute
47. 　　IV . Risk Warning
48. 　　The risk of AI industry commercialization failing to meet expectations . Currently, the commercialization model of AI products in various links is still in the exploratory stage. If the promotion pace of products in various links fails to meet expectations, it may have an adverse impact on the performance of related companies.
49. 　　Market competition increases risks . Overseas AI manufacturers have an advantage in the competition due to their first-mover advantage and strong technical accumulation. If the technological iteration of domestic AI manufacturers fails to meet expectations, their business conditions may be affected. At the same time, many domestic companies have invested in the research and development of AI products, and there may be a risk of homogeneous competition in the future, which will in turn affect the income of related companies.
50. 　　Policy uncertainty risk . The development of AI technology is directly affected by the policies and regulations of various countries. As AI penetrates various fields, the government may further introduce corresponding regulatory policies to regulate its development. If enterprises fail to adapt and comply with relevant policies in a timely manner, they may face corresponding penalties or even be forced to adjust their business strategies. In addition, policy uncertainty may also lead to errors in corporate strategic planning and investment decisions, increasing operational uncertainty .