DeepSeek was disruptive in January 2025.© Mojahid Mottakin/stock.adobe.com Headquarters:HangzhouRecent NewsFeb. 6, 2025, 10:20 PM UTC()House lawmakers push to ban AI app DeepSeek from US government devicesDeepSeek is a Chinese artificial intelligence (AI) company that rose to international prominence in January 2025 following the release of its mobile chatbot application and the large language model DeepSeek-R1. Released on January 10, it became the most downloaded app on Apple Inc.’s (AAPL) U.S. app store by January 27 and ranked among the top downloads on the Google Play store.

　　DeepSeek’s superior efficiency, affordability, and transparency compared to American AI companies led to a sharp decline in U.S. tech stocks on January 27. NVIDIA Corporation (NVDA) was particularly affected, with its share price plummeting 17% and losing nearly $600 billion in market capitalization—the largest one-day loss for a single company in U.S. stock market history. Many observers referred to the release of DeepSeek as a “Sputnik moment” that undermined widely held assumptions about American technological primacy.

　　DeepSeek-R1 is estimated to be 95% cheaper than OpenAI’s ChatGPT-o1 model and requires a tenth of the computing power of Llama 3.1 from Meta Platforms’ (META). Its efficiency was achieved through algorithmic innovations that optimize computing power, rather than U.S. companies’ approach of relying on massive data input and computational resources. DeepSeek further disrupted industry norms by adopting an open-source model, making it free to use, and publishing a comprehensive methodology report—rejecting the proprietary “black box” secrecy dominant among U.S. competitors.

　　DeepSeek AI models and chatbotsDeepSeek-V3: DeepSeek’s all-purpose chatbot assistant, positioned as a competitor to ChatGPT-4o. It was released in December 2024. It can respond to user prompts in natural language, answer questions across various academic and professional fields, and perform tasks such as writing, editing, coding, and data analysis. Like its American counterparts, it struggles with fact-checking, has a tendency to “hallucinate,” and often lacks deep insight, particularly in areas that require abstract thinking, such as beauty and humor.DeepSeek-R1: DeepSeek’s reasoning model, positioned as a competitor to ChatGPT-o1. It was released in January 2025. Reasoning models are designed for step-by-step logical deduction and complex problem-solving, making them better suited to high-level reasoning tasks than standard chatbot models. DeepSeek-R1 exemplifies the company’s emphasis on computational efficiency, achieving high-quality results with comparatively low resource requirements.Janus Pro: A multimodal AI model specializing in image generation and visual analysis, comparable to OpenAI’s DALL-E 3, Midjourney, and Stability AI’s Stable Diffusion. Available in 1B and 7B parameter versions, it can both generate high-quality images and engage in detailed discussions about real-world photographs. As visual understanding becomes an increasingly important frontier in AI, Janus Pro showcases DeepSeek’s capabilities in this segment, although it hasn’t been as disruptive as the company’s chatbot models.DeepSeek’s origins and rise in AIDeepSeek’s origins trace back to High-Flyer, a hedge fund cofounded by Liang Wenfeng in February 2016 that provides investment management services. Liang, a mathematics prodigy born in 1985 in Guangdong province, graduated from Zhejiang University with a focus on electronic information engineering. His early career centered on applying artificial intelligence to financial markets. By late 2017, most of High-Flyer’s trading activities were managed by AI systems, and the firm was well established as a leader in AI-driven stock trading.

　　Liang Wenfeng, founder and CEO of DeepSeek, delivers the keynote speech during the 10th China Private Equity Golden Bull Awards on August 30, 2019 in Shanghai, China.© Visual China Group/Getty ImagesAnticipating the growing importance of AI, Liang began accumulating NVIDIA graphics processing units (GPUs) in 2021, before the U.S. government placed restrictions on chip sales to China. This foresight enabled him to collect about 10,000 NVIDIA A100 GPUs, laying the groundwork for future AI endeavors.

　　In April 2023, High-Flyer announced the establishment of an artificial general intelligence lab dedicated to developing AI tools separate from its financial operations. By July 2023, this lab was incorporated as DeepSeek, with High-Flyer as its primary investor. Initially, venture capital firms were hesitant to fund DeepSeek because of uncertainties about its short-term profitability.

　　DeepSeek’s first breakthrough occurred in May 2024 with the release of the chatbot model DeepSeek-V2. This model gained immense popularity in China for its cost-efficiency, outperforming offerings from major tech companies such as ByteDance, Tencent, Baidu, and Alibaba. The success of DeepSeek-V2 triggered a price war, compelling each of these competitors to significantly cut prices on their AI models.

　　Building on this momentum, DeepSeek released DeepSeek-V3 in December 2024, followed by the DeepSeek-R1 reasoning model and its chatbot application in January 2025. These developments marked DeepSeek’s entry into the international market, challenging the prevailing assumption of U.S. dominance in AI. Shortly thereafter, Liang Wenfeng participated in a symposium with Chinese Premier Li Qiang, highlighting the government’s support for DeepSeek’s initiatives.

　　Unlike other Chinese technology companies, which are widely known for their “996” work culture (9 a.m. to 9 p.m., six days a week) and hierarchical structures, DeepSeek fosters a meritocratic environment. The company prioritizes technical competence over extensive work experience, often recruiting recent college graduates and individuals from diverse academic backgrounds. This approach emphasizes creativity, passion, and collaboration, drawing inspiration from Western work cultures.

　　Global bans and security breachesBy early February 2025, several governments and organizations had imposed restrictions on DeepSeek, citing national security and data privacy concerns. Taiwan was the first to take action, banning DeepSeek from all government agencies on January 27, with the Ministry of Digital Affairs warning that its use could “endanger national information security.”

　　The next day, Texas Governor Greg Abbott became the first U.S. official to restrict DeepSeek at the state level, prohibiting its use on government-issued devices. Soon after, the National Aeronautics and Space Administration (NASA) and the U.S. Navy issued internal bans, preventing employees from accessing DeepSeek services due to concerns about data vulnerabilities.

　　On January 30, Italy’s data protection authority, the Garante, blocked DeepSeek throughout the country, citing the company’s failure to provide adequate responses regarding its data privacy practices. On February 4, Australia banned DeepSeek from all government devices, with Home Affairs Minister Tony Burke emphasizing the security risks associated with foreign AI platforms.

　　Coinciding with increased scrutiny and regulatory actions, DeepSeek was targeted by a large-scale cyberattack, leading the company to suspend new user registrations outside mainland China on January 29.

　　The attack, which DeepSeek described as an “unprecedented surge of malicious activity,” exposed multiple vulnerabilities in the model, including a widely shared “jailbreak” exploit that allowed users to bypass safety restrictions and access system prompts. Further, a data breach led to the online leak of more than 1 million sensitive records, including internal developer notes and anonymized user interactions. The incident underscored both the security challenges facing AI platforms and the increasingly adversarial nature of the global race to dominate AI development.

　　Censorship in DeepSeek’s AI modelsAs of its January 2025 versions, DeepSeek enforces strict censorship aligned with Chinese government policies. It refuses to answer politically sensitive questions about topics including China’s top leader Xi Jinping, the 1989 Tiananmen Square incident, Tibet, Taiwan, and the persecution of Uyghurs.

　　When asked about these topics, DeepSeek either provides vague responses, avoids answering altogether, or reiterates official Chinese government positions—for example, stating that “Taiwan is an inalienable part of China’s territory.” These restrictions are embedded at both the training and application levels, making censorship difficult to remove even in open-source versions of the model.

　　American AI models also implement content moderation and have faced accusations of political bias, although in a fundamentally different way. Models such as ChatGPT, Claude, and Google Gemini are designed to prevent disinformation and minimize harm but have been observed to lean toward liberal political perspectives and avoid controversial topics. Unlike DeepSeek, which operates under government-mandated censorship, bias in American AI models is shaped by corporate policies, legal risks, and social norms.

　　Impact on U.S.–China AI competitionThe release of DeepSeek marked a paradigm shift in the technology race between the U.S. and China. Just weeks earlier, a short-lived TikTok ban in the U.S. had driven millions of American users to adopt the Chinese social media app Xiaohongshu (literal translation, “Little Red Book”; official translation, “RedNote”). The rapid rise of DeepSeek further demonstrated that Chinese companies were no longer just imitators of Western technology but formidable innovators in both AI and social media.

　　DeepSeek’s success also highlighted the limitations of U.S. semiconductor export controls. The Biden administration had imposed restrictions on NVIDIA’s most advanced chips, aiming to slow China’s development of cutting-edge AI. DeepSeek’s efficiency demonstrated that China possesses far more chips than was previously estimated, and has developed techniques to maximize computational power with unprecedented efficiency. This revelation raised concerns in Washington that existing export controls may be insufficient to curb China’s AI advancements.

　　Ethan Teekah