China’s artificial intelligence industry has found a new rising star in DeepSeek, a startup that has captured global attention despite being founded just three years ago. While the company has made remarkable strides in developing powerful AI models and disrupting the global AI market, its meteoric rise has brought significant challenges that threaten its future success.
From Startup to Spotlight: A Journey of Rapid Growth
DeepSeek was established in 2021 with a clear focus on AI research and development. The company began gaining serious market attention in the second half of 2023 when it started releasing its AI models to the public. The startup’s founder, Zhao Ming, is a former Google Brain researcher who brought his Silicon Valley experience to China with the ambitious goal of leading the country’s AI technological advancement.
From its inception, DeepSeek concentrated on developing large language models (LLMs) while adopting an unusual strategy in the AI industry: emphasizing both technical excellence and openness. Unlike most companies that keep their models proprietary, DeepSeek chose to release its core models as open-source, making them freely available to the global community.
Impressive Model Portfolio and Key Features
DeepSeek has developed several notable AI models, each with distinct capabilities:
DeepSeek-V2 represents the company’s expertise in Mixture-of-Experts (MoE) architecture, delivering economical training costs and efficient inference performance. This model successfully balances high performance with efficiency, making it suitable for various applications.
DeepSeek-Coder-V2 is specifically designed for programming language support, covering an impressive 338 programming languages. The model comes in two versions: a compact 16B parameter model and a large 236B parameter model. It incorporates proprietary attention mechanisms and MoE techniques to achieve superior performance in coding tasks.
DeepSeek-R1 is a massive model with 67.1 billion parameters, sharing structural similarities with GPT-4. This model excels particularly in mathematics and coding problem-solving while supporting multilingual data including English and Chinese. As a reinforcement learning-based reasoning model, it achieves both high-performance inference capabilities and cost efficiency.
DeepSeek-V3, announced in December 2024, has garnered significant attention for its ability to compete with Western-developed LLM models. The model demonstrates competitive performance against OpenAI’s models in mathematics, English, and coding while offering prices up to 95% lower, creating a global sensation in the AI community.
Facing Serious Challenges: A Crisis of Trust
Despite its achievements, DeepSeek currently faces multiple serious issues that have shaken user confidence and threaten its reputation.
1. Misinformation and Security Vulnerabilities
According to Fortune magazine and NewsGuard, a news reliability assessment organization, DeepSeek provides incorrect information in a staggering 83% of cases. Research teams have even warned that the system can be manipulated to generate dangerous content, including bomb-making instructions.
The security concerns extend beyond misinformation. Cybersecurity intelligence firm Kela has identified DeepSeek as particularly vulnerable to “jailbreak” attacks, where malicious actors can bypass the AI’s safety measures and prompt it to provide guidance on illegal activities such as money laundering or data theft malware creation.
2. Service Instability and Access Issues
DeepSeek has struggled with service instability due to sudden traffic surges, frequently overwhelming its API capacity. Performance degradation issues that began on January 27th continued through January 30th and 31st, with users regularly encountering “server busy, please try again later” messages.
However, some analysts suggest that DeepSeek’s access problems may extend beyond simple server overload, pointing to potential large-scale cyberattacks. These connectivity issues and service instability have negatively impacted DeepSeek’s overall performance and created significant user frustration.
3. Privacy and Data Protection Concerns
DeepSeek collects various types of personal information from users, storing this data on servers located in mainland China. This practice has raised significant concerns about data breaches and privacy protection.
Cybersecurity experts have discovered that user chat records and log data from DeepSeek’s database are publicly accessible, heightening privacy concerns. Multiple countries, including South Korea, are currently investigating DeepSeek’s data privacy practices. South Korea’s Personal Information Protection Commission plans to send an official inquiry to DeepSeek regarding its personal information collection items, procedures, and processing and storage methods.
4. International Usage Restrictions
While DeepSeek has gained attention for developing high-performance AI models at low cost using lower-specification semiconductors, its serious reliability and security vulnerabilities have led to potential international usage restrictions.
The U.S. Navy has completely banned the use of DeepSeek’s app, citing security vulnerabilities and concerns about Chinese government censorship. The White House is currently investigating DeepSeek’s potential impact on national security. Additionally, Taiwan’s Digital Ministry has prohibited public sector employees from using DeepSeek. These international restrictions have significantly damaged DeepSeek’s credibility.
5. Technical Limitations and Controversy
DeepSeek has recently become embroiled in allegations of unauthorized data appropriation. According to the Korea Economic Daily on January 31st, suspicions have been raised that DeepSeek illegally appropriated American AI technology, with Chinese government-affiliated commentators actively defending against these allegations.
While DeepSeek has achieved technological advancement despite U.S. AI chip export restrictions, demonstrating that AI technology development doesn’t necessarily require the latest hardware, questions remain about the validity of these claims. George Mason University Professor Tyler Cowen wrote in a January 9th Bloomberg column that AI technology doesn’t necessarily depend on the latest high-performance chips. However, Cowen noted that while he agrees with AI chip export restrictions for national security reasons, excessive measures could reduce their effectiveness. This raises concerns that countries like Russia, Iran, and Pakistan, which face AI chip supply restrictions due to U.S. sanctions, could develop similar programs.
Future Outlook and Challenges
DeepSeek’s emergence is expected to intensify competition in the AI market. The company’s AI models have demonstrated the potential to replace expensive GPU infrastructure, successfully maintaining performance while reducing costs by using H800 chips instead of NVIDIA’s H100. This development could enable small and medium-sized enterprises to easily adopt AI technology without requiring expensive infrastructure, potentially bringing significant changes across industries.
However, DeepSeek must address several challenges to sustain its future success. First, the company’s claims about using low-performance semiconductors require verification. According to NH Investment Securities researcher Lim Ji-yong, “Information about pre-training clusters has not been verified, so it would be premature to conclude that there is an advantage in terms of cost efficiency.”
Second, DeepSeek must improve its AI model stability. The company has experienced frequent disruptions, with “server busy” messages appearing every five minutes and operations stopping. This phenomenon, attributed to higher-than-expected global traffic, occurs more frequently and severely compared to ChatGPT’s early popularity period.
Finally, DeepSeek must strengthen its competitiveness in the global market. While the company’s AI models have succeeded in China, numerous competitors still exist in the global market. Therefore, DeepSeek needs to develop strategies to strengthen its position in the international arena.
As DeepSeek continues to navigate these challenges, its ability to address security concerns, improve service stability, and maintain its technological edge will determine whether it can fulfill its promise of democratizing AI technology or become another cautionary tale in the rapidly evolving AI landscape.
