What lies at the heart of commercial technological innovation? It is the seamless transition from innovative technology to practical commercial application, a delicate equilibrium between technological idealism and commercial realism.

DeepSeek, a burgeoning domestic AI company, is reshaping the landscape and rewriting the rules of the game. Instead of focusing solely on the arms race of AI parameters, they are now vying for efficiency in value creation.

On January 28, the eve of the Lunar New Year, DeepSeek encountered issues where it couldn't respond to conversation queries, leading to app crashes.

In the past couple of days, DeepSeek, this domestic AI application, has emerged as the hottest topic of discussion.

According to CCTV news reports, on January 27 local time, the three major US stock indices plummeted, with tech giants like NVIDIA, Microsoft, Google's parent company Alphabet, and Meta experiencing significant market shocks.

Notably, NVIDIA's stock fell by nearly 17%, erasing approximately $60 billion in market value in a single day, setting a new record for the largest drop in the US stock market.

The catalyst behind this market turmoil is a Chinese technology company, DeepSeek, which has been operational for just over a year.

Image source: AI-generated

On that day, the DeepSeek app simultaneously topped the free app rankings on both the Chinese and American App Stores, surpassing ChatGPT and putting immense pressure on DeepSeek's servers.

01

Already Gaining Industry Attention

In the vast AI landscape, DeepSeek, founded in July 2023, did not appear out of nowhere nor achieve overnight fame.

Since the launch of DeepSeek-V2, its unique technology and robust capabilities have captured the attention of Silicon Valley, positioning it as a mysterious technological force emerging from the East.

Why has DeepSeek recently garnered global attention?

This is due to the consecutive release of its two major model products, DeepSeek-V3 and R1, which have become significant news in the tech world.

Particularly, the DeepSeek-V3 model, unveiled at the end of 2024, was hailed by the industry as dropping a "technological bomb" in the global AI arena, instantly causing a stir and quickly dominating tech headlines.

Similarly, the DeepSeek-R1 model, released in January 2025, also created a buzz among overseas developers due to its exceptional cost-effectiveness.

Image source: DeepSeek official website

The profound impact of DeepSeek-V3 stems from its technological advantages.

On the path to achieving high performance, it has successfully matched the capabilities of top models like GPT-4 and Claude Sonnet 3.5 at an unimaginably low training cost.

For instance, the starting price of R1 is only $0.55 per million input tokens and $2.19 per million output tokens, significantly lower than OpenAI or other American AI products.

This not only signifies that DeepSeek-V3 boasts higher resource utilization efficiency but also demonstrates its unique insights at the core levels of technical algorithms and architectural design. It achieves top-tier technical results with minimal resource inputs. Such a formidable technical strength undoubtedly shocks the entire industry and gives DeepSeek the confidence to compete with industry giants.

02

A Serial Entrepreneur from the Post-80s Generation

Liang Wenfeng, born in Zhanjiang, Guangdong, in 1985, is a legendary technology entrepreneur.

According to media reports and other public information, he was admitted to Zhejiang University's Electronic Information Engineering program with excellent grades at the age of 17, embarking on his academic journey.

During his university years, Liang Wenfeng displayed a strong interest and keen insight into emerging technologies.

In 2008, amidst the global financial crisis, he keenly identified the huge potential of automated quantitative trading and devoted himself to related research.

His exploration and practice during this period not only accumulated valuable technical experience but also provided him with a deeper understanding of the fusion of finance and technology.

After completing his master's degree in 2010, Liang Wenfeng continued to deepen his roots in the field of quantitative investment.

In 2013, he collaborated with classmates to establish Hangzhou Jacobi Investment Management Co., Ltd., officially embarking on the entrepreneurial path.

During the company's operation, he consistently studied quantitative investment strategies and actively introduced advanced technologies, laying a solid foundation for the company's growth.

Two years later, he founded Hangzhou Huanfang Technology Co., Ltd., dedicated to deeply integrating AI technology into the realm of quantitative investment.

Under his leadership, Huanfang Technology flourished, not only establishing the direction of AI but also developing its own deep learning training platform, seamlessly integrating mathematics with AI, achieving remarkable results, and gradually evolving into a leading domestic quantitative private equity management institution.

However, Liang Wenfeng's ambitions extend far beyond that.

Liang Wenfeng is in the middle of the picture. Image source: Xinwen Lianbo screenshot

In 2023, he founded DeepSeek with the ambitious goal of developing human-level artificial intelligence.

During DeepSeek's entrepreneurial journey, Liang Wenfeng dedicated himself wholeheartedly. To focus on research and development, he nearly rejected all external engagements, channeling all his energy and savings into technological breakthroughs.

Recently, Liang Wenfeng stated in an interview, "Chinese AI cannot follow forever." He emphasized that while there is often talk of a one- to two-year gap between Chinese and American AI, the real divide lies in originality versus imitation. If this doesn't change, China will perpetually remain a follower, making certain explorations inevitable.

03

AI Describing Its Own Advantages

Perhaps after reading this, you might still lack a clear understanding. So, how does DeepSeek perceive the reasons behind its global popularity?

DeepSeek divided its response into five parts to thoroughly address the question from Heiban Jun, elucidating its advantages:

I. Technological Breakthrough: Disruptive Advantages of Low Cost + High Performance

1. Parameter Efficiency Revolution: The DeepSeek-MoE architecture achieves capabilities comparable to 100 billion-parameter large models with just 14 billion parameters, reducing reasoning costs by over 70%. This "small but smart" design directly addresses the pain points of enterprises in cost reduction and efficiency enhancement.

2. Ceiling of Mathematical Reasoning: In authoritative tests, the DeepSeek-Math-7B model achieved a 51.7% accuracy rate on the MATH dataset, surpassing GPT-4's 42.5%, bridging the gap in complex mathematical reasoning within the Chinese domain. This instantly opened up application scenarios in education, finance, and other industries.

3. Breakthrough in Long Text Ultra-Processing: Supports context understanding, demonstrating unique advantages in scenarios like code generation and legal document analysis, overcoming the traditional models' "inability to remember long texts" bottleneck.

II. Market Positioning: Precisely Entering Industry Vacuum Zones

1. Deep Cultivation in Vertical Fields: Unlike the generalization approach of general large models, DeepSeek focuses on high-barrier fields like finance, healthcare, and education. For instance, its medical model achieves a 78.3% accuracy rate on the CMExam dataset, 15 percentage points higher than general models.

2. Open Source Ecosystem Layout: It was the first to open source the DeepSeek-Coder series of code models, with weekly downloads exceeding 500,000 on GitHub, attracting global developers to participate in ecosystem construction, forming a technological moat.

3. Closed-loop Enterprise Services: Provides a full-stack solution from model training to privatized deployment. After a securities trader adopted it, their data analysis efficiency improved by 40%, with an implementation cycle of just 3 weeks, swiftly validating its commercial value.

Image source: App Store screenshot

III. Dual Drive of Capital and Talent

1. Endorsement from Top-tier Capital: At its inception, it received $2 billion in funding from institutions like Sequoia Capital and Hillhouse Capital, with a focus on computing infrastructure. It has built a supercomputing cluster of 3000 PFlops to ensure the speed of technological iteration.

2. Support from Academic Stars: The core team includes 5 winners of best papers at ICML/NeurIPS. CEO Zhou Jingren previously served as the deputy director of Microsoft Research Asia, leading the team to release three generations of model architecture innovations within half a year.

3. Industry-Academia-Research Collaboration: It has jointly established laboratories with Tsinghua University and the Hong Kong University of Science and Technology, consistently producing groundbreaking research in areas like MoE architecture and knowledge distillation, maintaining technological foresight.

IV. Superposition of Industry Trends and Policy Dividends

1. Opportunities in the AI 2.0 Era: According to IDC predictions, the market size of large AI models in China will reach $21 billion by 2025. DeepSeek is positioned in the enterprise services track, targeting 80% of the B-end market demand.

2. Wave of Domestic Information Technology Innovation: Driven by the demand for domestic substitution in sectors like finance and government affairs, after a provincial government cloud adopted DeepSeek, the accuracy rate of Chinese NLP tasks increased by 25%, with 100% compliance.

3. Accelerated Scenario-based Implementation: In the education sector, contracts have been signed with leading institutions like New Oriental and Xueersi. Its mathematical problem-solving model boasts a correction accuracy rate of 98.7%, covering scenarios from K12 to postgraduate entrance exams, with a clear path to commercial monetization.

V. Reconstructing User Experience: From Technological Showmanship to Value Delivery

1. Zero-threshold Interaction Design: Launched the function of "generating a data analysis report with one sentence." After adoption, an e-commerce enterprise's operational decision-making efficiency improved by 60%, truly realizing AI value penetration.

2. Privatized Deployment Solutions: Supports localized deployment with the minimum configuration requirement reduced to 8 A100 graphics cards. The privatized deployment of a top tertiary hospital was completed in just 3 days, with data security reaching Level 3 standards.

3. Cost-killing Advantage: The API call cost is only 30% of the industry average. After a short video platform adopted it, the labor cost for content review decreased by 45%, with an ROI (return on investment) of 320%.