- 30
- March
DeepSeek Series EP.1 — Since early 2025, the global AI industry has been rocked by DeepSeek, a Chinese AI company that was virtually unknown before it built AI models rivaling GPT-4 on a fraction of the budget. The shockwave was so severe that NVIDIA stock plunged $600 billion in a single day, leading media outlets worldwide to call it the "Sputnik Moment" of AI. This article is EP.1 of our DeepSeek Series, giving you a comprehensive introduction to DeepSeek — who they are, where they came from, what their models can do, and whether Thai organizations should consider using them.
Quick Summary — What is DeepSeek?
- DeepSeek is a Chinese AI company founded in July 2023 by Liang Wenfeng, owner of the High-Flyer hedge fund
- Key models: DeepSeek-V3 (671B parameters) and DeepSeek-R1 (Reasoning model)
- Open-source under MIT License — free to download and run on your own servers
- API pricing is 10-50x cheaper than GPT
- V3 was trained on a budget of only $5.6 million (GPT-4 cost $100 million+)
- Reached #1 on the US App Store (January 2025)
Who is DeepSeek? — Origins and Founder
DeepSeek (Chinese name: 深度求索) is an AI company founded in July 2023 by Liang Wenfeng (梁文锋), a Chinese entrepreneur in his early 40s who owns High-Flyer (幻方量化), a quantitative hedge fund that uses AI and algorithmic trading to manage assets worth over $8 billion.
DeepSeek is headquartered in Hangzhou, Zhejiang Province, China. Liang Wenfeng holds approximately 84% of DeepSeek through multiple shell corporations, making him the company's de facto controller.
The origins of DeepSeek trace back to High-Flyer establishing an internal AGI Lab in April 2023 to research Artificial General Intelligence (AGI). This lab was spun off as an independent company in July of the same year. Notably, Liang has been clear that DeepSeek is not focused on short-term revenue — its long-term goal is to build AGI (Artificial General Intelligence), an AI that can think and reason like a human. This is the same objective pursued by OpenAI and other leading AI companies.
Timeline: From Zero to World Contender
In just two years, DeepSeek went from an unknown company to one of the most important players in the global AI landscape. Here are the key milestones:
| Date | Event | Significance |
|---|---|---|
| Apr 2023 | AGI Lab established within High-Flyer | Beginning of AI research |
| Jul 2023 | DeepSeek officially founded | Spun off from High-Flyer as independent company |
| Nov 2023 | Launch of DeepSeek Coder, LLM, Chat | First model family — began gaining recognition |
| May 2024 | DeepSeek-V2 released | Gained fame in China for extremely low pricing, triggering an "AI price war" |
| Dec 2024 | DeepSeek-V3 released | 671B parameters, MoE architecture, trained for only $5.6M |
| Jan 2025 | DeepSeek-R1 + Chatbot launch | Reasoning model, hit #1 on US App Store, NVIDIA lost $600B in market cap |
| Mar 2025 | DeepSeek-V3-0324 (update) | Performance improvements to V3 |
| May 2025 | DeepSeek-R1-0528 (update) | Enhanced reasoning capabilities |
| Aug 2025 | DeepSeek-V3.1 (Hybrid V3+R1) | Combined V3 and R1 capabilities into a single model |
| Nov 2025 | DeepSeek-V3.2 (Sparse Attention) | New technique for memory savings and faster processing |
| 2026 | DeepSeek V4 expected | Not yet released as of March 2026 — rumored to be a competitor to GPT-5.4 |
Why Did DeepSeek Explode in Popularity?
DeepSeek is not the first Chinese AI company, but it is the first to make Silicon Valley sit up and take Chinese AI seriously. Four key factors fueled its meteoric rise:
1. Radically Low Cost
DeepSeek trained its V3 model (671 billion parameters) on a budget of just $5.6 million, while OpenAI's GPT-4 reportedly cost $100 million or more. Furthermore, DeepSeek's API pricing is 10-50x cheaper than GPT — input tokens cost just $0.10-$0.55 per million tokens, compared to GPT's $2.00-$15.00. This raised a fundamental question: "If world-class AI can be built this cheaply, why are American AI companies spending billions?"
2. Fully Open-Source Under MIT License
DeepSeek chose to release all of its model weights under the MIT License, the most permissive open-source license available. Anyone can download, modify, or use DeepSeek models for commercial purposes — completely free. This stands in stark contrast to ChatGPT and Claude, both of which are closed-source. The open-source approach won DeepSeek a massive following among developers worldwide, especially among organizations that need to maintain full control over data security by running models on their own servers.
3. Built Despite US Sanctions
The United States banned exports of high-end AI chips (such as NVIDIA H100 and A100) to China starting in 2022. Despite this, DeepSeek managed to build world-class models using older, restricted chips (NVIDIA H800, a deliberately downgraded version for the Chinese market). They achieved this through innovative techniques like Mixture of Experts (MoE) architecture and FP8 Mixed Precision Training to dramatically reduce computational requirements. This proved that sanctions may not be as effective at slowing Chinese AI development as intended.
4. The "Sputnik Moment" — A Stock Market Earthquake
On January 27, 2025, following the launch of DeepSeek-R1 and its rise to #1 on the US App Store, NVIDIA's stock plunged by $593 billion in a single day — the largest single-day loss in US stock market history. Media outlets around the world compared the event to a "Sputnik Moment" — a reference to when the Soviet Union launched the first satellite into space ahead of the United States in 1957, triggering the Space Race. This time, it marked the beginning of a genuine "AI arms race between superpowers", with far-reaching implications for AI Governance and technology policy worldwide.
Comparison Table — DeepSeek vs ChatGPT vs Claude
To provide a clear picture, here is a side-by-side comparison of DeepSeek with other leading AI platforms commonly used by Thai organizations:
| Feature | DeepSeek | ChatGPT | Claude |
|---|---|---|---|
| Company | DeepSeek (China) | OpenAI (USA) | Anthropic (USA) |
| Latest Model | V3.2 / R1 | GPT-5.4 | Claude Opus 4.6 |
| Open-source | MIT License | Closed | Closed |
| API Price (Input/1M tokens) | $0.10-$0.55 | $2.00-$15.00 | $3.00-$15.00 |
| Reasoning | R1 (Chain-of-Thought) | GPT-5.4 Thinking | Claude Extended Thinking |
| Data Storage | China | USA | USA |
| Self-hosting | Yes | No | No |
The table makes DeepSeek's two key strengths abundantly clear: low cost and open-source availability. However, the data storage location is a significant concern. For a deeper technical comparison with the latest models, see our AI Model Comparison article.
DeepSeek's Core Models — A Closer Look
DeepSeek-V3 — The Flagship Language Model
DeepSeek-V3 is the company's flagship Large Language Model (LLM) with 671 billion parameters, built on the Mixture of Experts (MoE) architecture. For each token processed, only 32 billion parameters (out of 671B total) are activated, making it far more energy-efficient and faster than dense models of comparable size. V3 excels across a wide range of tasks including writing, summarization, analysis, and code generation.
DeepSeek-R1 — The "Think Before You Answer" Reasoning Model
DeepSeek-R1 is a reasoning model trained using Reinforcement Learning (RL) to "think step by step" before answering questions. R1 displays its thinking process to users, showing how it arrives at its conclusions — similar to GPT-5.4 Thinking or Claude Extended Thinking. What made headlines in the AI research community is DeepSeek's discovery that reasoning can "emerge naturally" from RL training without requiring explicit supervised instruction — a groundbreaking finding that generated significant excitement among AI researchers.
DeepSeek Coder — Specialized for Programming
DeepSeek Coder is a specialized model for software development, supporting over 80 programming languages. It achieves performance comparable to GPT-4 on major coding benchmarks, making it an excellent choice for developers who need AI-assisted coding capabilities.
DeepSeek-V3.1 / V3.2 — Next-Generation Upgrades
DeepSeek-V3.1 (August 2025) is a hybrid model that combines the capabilities of V3 and R1 — it can engage in deep reasoning when needed while still responding quickly to straightforward questions. V3.2 (November 2025) introduced Sparse Attention, a technique that reduces memory usage and speeds up processing — aligning with the same trends driving the development of Agentic AI.
Risks You Need to Know
Key Risks of DeepSeek to Consider:
- Data stored in China: If you use DeepSeek's API (without self-hosting), all data is sent to servers in China, which are subject to China's National Intelligence Law of 2017 — the Chinese government can legally request access to any data stored on domestic servers.
- Banned in multiple countries: Italy banned it over GDPR concerns; Australia banned it on government devices; Taiwan prohibited its use in government agencies; the US Navy banned it from all devices.
- Security vulnerabilities: Wiz Research discovered that DeepSeek had a publicly accessible database containing over 1 million records, including chat histories, API keys, and system logs. While the issue was resolved, it raised serious questions about the company's cybersecurity standards.
- Censorship: DeepSeek refuses to answer questions that the Chinese government censors, including topics like Tiananmen Square, Taiwan's status, and Tibet. This can affect the reliability of its outputs for certain types of research.
- Service disruptions: DeepSeek has experienced outages lasting up to 7 hours due to DDoS attacks, posing a risk for organizations that depend on API availability.
DeepSeek and Thai Organizations — Should You Use It?
The short answer is yes, but with caution — particularly regarding Data Privacy and Thailand's PDPA (Personal Data Protection Act). For Thai organizations considering DeepSeek, here is a breakdown by use case:
Suitable for:
- Internal tasks that do not involve sensitive data — e.g., summarizing articles, translation, general-purpose coding
- Experimenting with AI at low cost, ideal for SMEs with limited budgets
- Self-hosting — downloading the model and running it on your organization's own servers so that no data leaves your premises
- R&D and prototyping work that has not yet reached production
Not suitable for:
- Processing customer data or personal information (potential PDPA violation)
- Financial data, financial statements, or transaction records
- Organizations with strict compliance requirements — e.g., banks, insurance companies, government agencies
- Tasks requiring 100% accuracy, especially in legal and medical contexts
For organizations that want AI to enhance their ERP systems — whether for data analysis, report generation, or AI-assisted accounting — you should choose an AI provider with clear compliance policies, in-country data storage, or at minimum a transparent Data Processing Agreement (DPA). As of now, DeepSeek does not offer a DPA comparable to what OpenAI or Anthropic provide.
DeepSeek Series — Read More
DeepSeek Series — 5 Episodes on the Chinese AI Challenger:
- EP.1: What is DeepSeek? — The Chinese AI That Shook the World (this article)
- EP.2: Mixture of Experts — The Technique That Makes It 10x Cheaper
- EP.3: Risks of Chinese AI — What Thai Organizations Must Know
- EP.4: Running DeepSeek On-Premise — Is It Worth It?
- EP.5: Can DeepSeek Really Help with ERP?
DeepSeek has proven that world-class AI does not have to be expensive — but the real question for Thai organizations is: "Is the low price worth the risk?"
— Saeree ERP Team
