In the rapidly evolving landscape of artificial intelligence (AI), China’s DeepSeek AI chatbot has emerged as a formidable contender. It was established in mid-2023 by Liang Wenfeng the co-founder of High-Flyer Capital Management, a quantitative hedge fund or quant fund in Hangzhou, Zhejiang. In early 2025, DeepSeek AI Chatbot gained global attention for launching an extremely efficient and inexpensive generative AI model. This launch triggered significant market shifts, privacy concerns, and geopolitical debates, making it a key player in the global AI race.
Within 18 months from its founding, a multidisciplinary team of developers and AI researchers at DeepSeek developed and deployed DeepSeek-R1 and DeepSeek-V3, a generative AI model that is capable of reasoning at some level, coding, understanding natural language, and generally generating content at par with open AI’s GPT-4 or Meta’s Llama 2.
DeepSeek-V3 is a powerful, general-purpose mixture-of-experts (MoE) model, while DeepSeek-R1 is a specialised version of V3 that uses reinforcement learning to focus on advanced and transparent reasoning for speed and efficiency.
Efficient and Cost-Effective Architecture
DeepSeek's unique innovation lies in its use of the MoE model coupled with multi-token processing. MoE architectures dynamically allocate subsets of the general model parameters, or ‘experts,’ to different inputs, such that only a few relevant experts activate for each prompt. This reduces the computational and memory requirement to a drastic level, with DeepSeek allegedly utilising 75 per cent less graphics processing unit (GPU) memory during inference when compared to other similar-scale models.
The startup trained its model using an estimated 2,000 Nvidia H800 GPUs (chips) at an expenditure of around US$ 5.6 million— far less compared to how much OpenAI actually spent, estimated somewhere in the hundreds of millions. Deepseek-V3’s full training required 2.788 million H800 GPU hours and covered 14.8 trillion tokens, utilising novel techniques like multi-head Latent Attention and auxiliary-loss-free load balancing. (H800 is the world’s most advanced chip ever build-with 80 billion transistors. The US has banned export of H800 to China.)
Features and User Experience
DeepSeek open approach aids researchers, developers, and startups all over the world without the barrier of costly licences.
DeepSeek provides a chatbot interface to conversational AI-based tasks such as answering complex questions, reasoning, coding, understanding natural language, debugging, summarisation, translation, creative writing, content creation, and making logical puzzles and games.
Open Weight Licensing and Accessibility
Unlike many commercial AI models that keep weights proprietary, DeepSeek released its model weights under a permissive licence that is close to that of the Massachusetts Institute of Technology (MIT). Independent security research by Wiz Research revealed a misconfigured clickHouse database exposing over 8 million user records and sensitive API keys on the public Internet, including chat logs, IP addresses, operating system information, and user interactions. Researchers also flagged additional vulnerabilities—outdated encryption, SQL injection risks, and weak app transport security—alongside a 91 per cent jailbreaking and 86 per cent promt-injection failure rate and emerging phishing schemes targeting users.
Jailbreaking is a technique to bypass the safety features of a large language model (LLM), allowing it to generate restricted content, normally trained to avoid.
In promt-injection attacks, hackers manipulate generative AI systems by feeding them harmful inputs disguised as legitimate prompts.
Within its first 18 days, DeepSeek became the number one free app on the Apple App Store in the USA, with 16 million downloads—surpassing ChatGPT’s 9 million during the same period.
Privacy, Security, and Content Moderation
Data Collection and Exposure Risks DeepSeek has faced intense scrutiny over its data handling practices. DeepSeek also reportedly collects extensive telemetry data, with information stored on servers physically located in China, which raises concerns about government access and compliance with local laws requiring cooperation with state intelligence.
Content Censorship and Self-Moderation Following Chinese government mandates, DeepSeek’s chatbot restricts discussions on sensitive political topics including Taiwan, Xinjiang, Tibet, and human rights issues. Such content restrictions have led to criticism from freedom of speech advocates and raised questions about the role of AI in reinforcing state narratives.
The company employs automated filters and prompt engineering techniques to ensure compliance with the state public opinion guidance rules, limiting free expression on certain subjects. Many governments have barred DeepSeek on federal devices, citing risks of data harvesting and potential influence by the Chinese Communist Party:
- Federal and state governments have banned DeepSeek’s use due to privacy, security, and geopolitical concerns.
- In Czech Republic, the government has prohibited the use of DeepSeek within public administration over data security and sovereignty issues. There The government barred DeepSeek on federal devices, citing risks of data harvesting and potential influence by the Chinese Communist Party.
- In Australia, the majority of corporations and government bodies restricted DeepSeek citing unacceptable security and compliance risks.
- In Japan, DeepSeek is banned on official government devices to protect data confidentiality.
- In South Korea, the Personal Information Protection Commission suspended new downloads of DeepSeek and restricted its usage within government agencies.
Geopolitical Implications
DeepSeek has intensified the AI rivalry between China and the USA. It represents a strategic effort by China to assert technological leadership amid USA. Its emergence has triggered a Sputnik moment for the AI industry, sparking increased investment in AI infrastructure and innovation in the West. Chinese Premier Li Qiang publicly called for the establishment of a global AI cooperation organisation to promote shared governance and reduce fragmented regulation, sanctions, and export restrictions.
Accuracy and Reliability Challenges
Despite its popularity, DeepSeek has faced criticism for accuracy and misinformation. For instance, in misinformation audits, it failed to reliably debunk false claims in 83 per cent of cases, ranking near the bottom among 11 AI systems tested. This raises concerns about this chatbot’s suitability for critical applications, especially in education, journalism, and healthcare sector. This has significant implications for:
- Small and medium enterprises (SMEs) in developing countries seeking AI capabilities,
- Democratising AI access beyond well-funded Silicon Valley firms,
- Challenging the narrative that AI development requires massive capital expenditure.
By releasing an open weight model, it enables innovation and customisation by researchers worldwide, potentially accelerating AI adoption in diverse sectors.
An open weight model can be downloaded, used, and five-tuned without need access to the original training data or code. It allows transparency for community-driven research and development.
Applications in Health Care
DeepSeek has been adopted by many of China’s tertiary hospitals to support clinical workflows to assist in diagnostics, analyse medical imaging, streamline patient data management, and support in decision-making.
While showing efficiency gains, the deployment also highlights ethical and regulatory challenges—including automation bias, data privacy risks, and legal liability if AI advice leads to medical errors.
Economic and Innovation Impact
DeepSeek is an affordable and cost-efficient AI model which lowers barriers to entry for AI research and development globally. It acts like a driver for rethinking AI infrastructure investment, regulation, and ethical frameworks and putting China in line with latest AI development models. However, experts caution about the Jevons paradox: improved efficiency can lead to increased overall consumption as AI adoption grows, potentially raising total energy use despite localised savings.
Jevons Paradox, first proposed by William Stanley Jevons in 1865, is an economic principle which suggests that improvements in the efficiency of using a resource tend to increase its overall consumption, rather than reduce it.
Continuing Developments: DeepSeek-V3.1
On August 21, 2025, DeepSeek unveiled version 3.1—an upgraded model featuring hybrid inference, faster processing, and enhanced agent capabilities. API pricing will be updated starting in September, 2025. DeepSeek-V3.1 is a 685-billion-parameter model quietly released via Hugging Face, a platform where users can upload, find and down-load pre-trained AI models. It scored 71.6 per cent on the Aider coding benchmark (or pass rate), slightly above Claude Opus 4 and at a fraction of cost—estimated at just US$ 1.01 per coding task vs approximately US$ 70 for competitors. It handles up to 128,000 tokens of context, supports BF16, F8_E4M3, and F32 formats, and integrated novel thinking tokens and web search capabilities—all while being fully open-source.
Conclusion
DeepSeek represents a paradigm shift in AI development and deployments—offering cost-effective, open-weight generative AI that competes with western systems, disrupting markets, redefining investment pressures, and raising complex privacy, security, and governance questions.
Its rise underscores the urgency for transparent, global dialogue to ensure AI benefits society while mitigating risks. As AI technologies proliferate, balancing innovation, privacy, security, and ethical considerations remains critical.
© Spectrum Books Pvt Ltd.
