HTLJ
What is DeepSeek?
In December 2024, Chinese startup DeepSeek shook up the AI landscape with its open-source model V3, claiming to match U.S. competitors while requiring far less computing power.[1] Led by Liang Wenfeng, the company quickly emerged as a challenger to OpenAI’s ChatGPT and Google’s Gemini.[2] When DeepSeek released its reasoning-focused R1 model on January 20—specialized for math, coding, and logical analysis—it shot to the top of Apple’s App Store charts in the U.S.[3] and garnered over 16 million downloads in 18 days.[4] This is twice the number of downloads ChatGPT received in the same period.[5]
To understand DeepSeek’s claimed breakthrough, it helps to understand how AI models are powered. Large Language Models, like DeepSeek and ChatGPT, rely on massive amounts of data and specialized computer chips called Graphics Processing Units (GPUs).[6] Originally designed for rendering video games, GPUs are now essential for AI training.[7] U.S. company Nvidia dominates the GPU market, holding 92% of the market share.[8] Its chips excel at parallel processing, allowing them to perform many calculations simultaneously, making them perfect for AI development.[9] However, these specialized chips are expensive and energy-intensive, with high-end Nvidia GPUs costing thousands of dollars each.[10] While the number of GPUs needed varies greatly depending on the complexity of the AI model and dataset size, training advanced models often requires multiple high-performance GPUs, leading to significant overall costs.[11]
What Makes It So Special?
What sets DeepSeek apart from other AI models is its claimed efficiency in resource usage. They built their technology using only about $6 million in raw computing power.[12] That’s roughly one-tenth of what Meta spent on its latest AI technology[13] and far less than the $100 million used to train ChatGPT’s 4o model.[14] DeepSeek also claims that its model requires just 2,000 GPUs, whereas leading models typically need around 16,000.[15]
However, these claims have faced skepticism from industry figures. Elon Musk, for instance, has questioned the company’s statements about its GPU usage, suggesting it actually possesses around 50,000 Nvidia chips.[16]
Beyond its resource efficiency claims, DeepSeek R1 distinguishes itself through its “thinking out loud” approach using Chain of Thought (CoT).[17] This step-by-step reasoning makes AI more transparent, accurate, and easier to debug.[18] This approach differs from ChatGPT’s black-box model, making DeepSeek R1 a more transparent decision-making model.[19]
Finally, DeepSeek’s decision to “open source” its model is also significant, as it allows other businesses and researchers to build upon and distribute their own products using the same technologies.[20] Microsoft Azure and AWS have already begun offering China’s model on their own servers, enabling local deployment.[21]
The combination of DeepSeek’s claimed cost efficiency, transparent reasoning through CoT, and open-source accessibility creates a compelling alternative in the AI landscape.
Why Does It Matter?
The arrival of DeepSeek has rattled Silicon Valley and reshaped expectations about AI development costs. Tech venture capitalist Marc Andreessen called it “AI’s Sputnik moment,” highlighting its potential significance in the global AI race.[22] President Trump also dubbed it a “wake up call” for U.S. tech companies.[23]
The impact has been particularly dramatic for Nvidia. Following DeepSeek’s December announcements, Nvidia experienced the biggest one-day loss in US market history, shedding almost $600 billion in market value and dropping from the world’s most valuable company to third place behind Apple and Microsoft.[24] Other companies such as OpenAI, Anthropic, and xAI also feel pressure to innovate rapidly following the market’s reassessment of AI technologies and the potential disruption caused by DeepSeek.[25]
However, DeepSeek struggles with content moderation. When tested by BBC, the AI showed clear limitations, suggesting Chinese government censorship.[26] When asked, “What happened at Tiananmen Square?” it simply responded, “Sorry, that’s beyond my current scope. Let’s talk about something else.”[27] This response exemplifies the broader phenomenon of the “splinternet”—the increasing fragmentation of the internet into distinct regional networks with varying rules and restrictions.[28] Such adherence to Chinese censorship requirements could significantly impede DeepSeek’s international expansion, especially in markets that value unrestricted access to information.[29]
The implications extend beyond just market valuations and censorship concerns. DeepSeek’s claimed breakthrough in efficient AI development could fundamentally reshape how AI models are built and deployed. If their approach proves successful, it could democratize AI development by making it more affordable for smaller companies and researchers. Lower barriers to research, development, and market entry, like cost, will likely allow for a more diverse range of players to experiment with AI innovation.
What are Its Legal Implications?
DeepSeek’s growing popularity brings important privacy concerns about how it handles and stores user data.[30] The company operates from China, and its privacy policy is clear, stating: “We store the information we collect in secure servers located in the People’s Republic of China.”[31] This statement worries privacy experts, as Chinese laws grant government authorities broad access to data stored within the country.[32]
Security expert Lukasz Olejnik from King’s College London warns against sharing personal information with the platform.[33] However, he notes that users can still use the model without sending their info to China; they can download DeepSeek’s open-source models and run them on their own computers.[34] This means users can use the technology without sending their data to DeepSeek’s servers in China. However, if users input data from the app or the website, that information will be stored on DeepSeek’s servers in China, where it may be subject to Chinese data regulations, which may allow government access or impose restrictions based on local laws.[35]
Moreover, DeepSeek seems particularly vulnerable to various attempts to weaponize the tool for harm. Security researchers from Cisco and the University of Pennsylvania are publishing findings showing that, when tested with 50 malicious prompts designed to elicit toxic content, DeepSeek’s model “did not detect or block a single one.”[36] This failure to filter attacks not only raises serious ethical questions about the responsible development of AI but also exposes the company to potential legal liabilities, like regulatory fines and lawsuits.
Export controls pose an additional challenge, as U.S. restrictions limit Chinese companies’ access to advanced AI chips, potentially affecting DeepSeek’s development capabilities.[37] Intellectual property questions also arise from DeepSeek’s open-source approach. While sharing code freely promotes innovation, it raises questions about liability when others modify their technology.[38]
DeepSeek must also comply with different AI regulations worldwide.[39] The EU’s AI Act, for example, imposes strict rules on AI systems.[40] The Act classifies systems based on risk into four tiers: unacceptable, high, limited, and minimal, with the strictest controls imposed on those that could significantly impact safety or fundamental rights.[41] Due to concerns about censorship and hidden vulnerabilities, DeepSeek may face greater scrutiny, especially if regulators suspect its outputs are influenced by state-controlled narratives.[42] There are also concerns about whether DeepSeek’s training data meets international cybersecurity and bias management standards.[43]
Finally, DeepSeek’s terms require all disputes to be handled under Chinese law in Chinese courts.[44] If negotiations fail, cases must be filed in Hangzhou where DeepSeek is registered.[45] This creates significant challenges for foreign users, as Chinese courts don’t allow pre-trial discovery and most foreign companies lack experience with Chinese litigation.[46] In practice, this means users have limited legal options if DeepSeek violates its terms.[47]
What’s Next?
DeepSeek’s future impact on AI is still unfolding. As a private company, investors cannot buy shares on public exchanges. This limits investment options but may give DeepSeek more freedom to develop. The AI competition continues to intensify. OpenAI has already announced a more powerful system than DeepSeek’s R1, however, this system remains unreleased to the public. The race for AI supremacy shows no signs of slowing down.
[1] Cade Metz, What to Know About DeepSeek and How It Is Upending A.I., N.Y. Times (Jan. 27, 2025), https://www.nytimes.com/2025/01/27/technology/what-is-deepseek-china-ai.html.
[2] Kelly Ng et. al., DeepSeek: The Chinese AI app that has the world talking, BBC News (Feb. 3, 2025), https://www.bbc.com/news/articles/c5yv5976z9po; Aimee Picchi, What is DeepSeek, and why is it causing Nvidia and other stocks to slump?, CBS News (Jan. 28, 2025, 12:29 PM), https://www.cbsnews.com/news/what-is-deepseek-ai-china-stock-nvidia-nvda-asml/.
[3] Picchi, supra note 2.
[4] Muslim Farooque, DeepSeek AI Tops App Downloads, Surpasses ChatGPT but Faces Global Scrutiny, Yahoo! Fin. (Feb. 4, 2025), https://finance.yahoo.com/news/deepseek-ai-tops-app-downloads-141941732.html.
[5] Id.
[6] Rick Merritt, Why GPUs Are Great for AI, Nvidia Blog (Dec. 4, 2023), https://blogs.nvidia.com/blog/why-gpus-are-great-for-ai/; Matt Sarrel, CPU vs. GPU: What’s best for machine learning, Aerospike (Oct. 17, 2024), https://aerospike.com/blog/cpu-vs-gpu/.
[7] Karen Freifeld, How the new AI chip rule from the US will work, Reuters (Jan. 13, 2025), https://www.reuters.com/technology/artificial-intelligence/how-new-ai-chip-rule-us-will-work-2025-01-13/.
[8] Joaquin Fernandez, The leading generative AI companies, IOT Analytics (Mar. 4, 2025), https://iot-analytics.com/leading-generative-ai-companies/.
[9] Merrit, supra note 6.
[10] Stewart Bendle, GPU Price Index 2025: Lowest price on every graphics card from Nvidia, AMD, and Intel today, Tom’s Hardware (Jan. 27, 2025), https://www.tomshardware.com/news/lowest-gpu-prices.
[11] Hardware Recommendations for Machine Learning/ AI, Puget Sys., https://www.pugetsystems.com/solutions/ai-and-hpc-workstations/machine-learning-ai/hardware-recommendations/ (last visited Mar. 16, 2025).
[12] Metz, supra note 1.
[13] Id.
[14] Will Knight, OpenAI’s CEO Says the Age of Giant AI Models Is Already Over, Wired (Apr. 17, 2023), https://www.wired.com/story/openai-ceo-sam-altman-the-age-of-giant-ai-models-is-already-over/.
[15] Metz, supra note 1.
[16] Elon Musk not convinced with China’s DeepSeek that rattled the US tech market. His one word reply is the proof, The Econ. Times, https://economictimes.indiatimes.com/news/international/global-trends/us-news-elon-musk-deepseek-ai-elon-musk-not-convinced-with-chinas-deepseek-that-rattled-the-us-tech-market-his-one-word-reply-is-the-proof/articleshow/117678549.cms (last updated Jan. 28, 2025, 11:24 AM).
[17] Tobias Mann, China’s Deepseek just emitted a free challenger to OpenAi’s o1- here’s how to use it on your PC, The Register (Jan. 26, 2025), https://www.theregister.com/2025/01/26/deepseek_r1_ai_cot/.
[18] Id.
[19] Id.
[20] Metz, supra note 1.
[21] Channy Yun, DeepSeek models now available on AWS, AWS News Blog (Jan. 30, 2025) https://aws.amazon.com/blogs/aws/deepseek-r1-models-now-available-on-aws; Asha Sharma, DeepSeek R1 is now available on Azure AI Foundry and GitHub, Microsoft Azure (Jan. 29, 2025), https://azure.microsoft.com/en-us/blog/deepseek-r1-is-now-available-on-azure-ai-foundry-and-github/.
[22] Metz, supra note 1.
[23] David Ingram, Trump Says China’s DeepSeek AI ‘should be a wake up call’ for American tech companies, NBC News (Jan. 27, 2025), https://www.nbcnews.com/tech/innovation/trump-china-deepseek-ai-wake-call-rcna189526.
[24] Laura Bratton, Nvidia stock begins recovery after DeepSeek AI frenzy prompted near $600 billion loss, Yahoo! Fin. (Jan. 28, 2025), https://finance.yahoo.com/news/nvidia-stock-begins-recovery-after-deepseek-ai-frenzy-prompted-near-600-billion-loss-134240328.html.
[25] Id.
[26] James T. Areddy, DeepSeek’s Chatbot Works Like Its U.S. Rivals- Unitl You Ask About Tianamen, The Wall St. J. (Jan. 30, 2025), https://www.wsj.com/tech/ai/deepseek-chatgpt-tiananmen-square-efcd9938?.
[27] Id.
[28] Eric Goldman, Internet Law: Cases & Materials 8 (2024).
[29] Areddy, supra note 26.
[30] Matt Burgess et al., DeepSeek’s Popular AI App is Explicitly Sending US Data to China, Wired (Jan. 27, 2025), https://www.wired.com/story/deepseek-ai-china-privacy-data/?.
[31] Id.
[32] Id.
[33] Id.
[34] Id.
[35] Id.
[36] Matt Burgess & Lily H. Newman, DeepSeek’s Safety Guardrails Failed Every Test Researchers Threw at Its AI Chatbot, Wired (Jan. 31, 2024), https://www.wired.com/story/deepseeks-ai-jailbreak-prompt-injection-attacks/.
[37] Karen Freifeld, US looking into whether DeepSeek used restricted AI chips, source says, Reuters (Jan. 31, 2025), https://www.reuters.com/technology/us-looking-into-whether-deepseek-used-restricted-ai-chips-source-says-2025-01-31/.
[38] Robert Hulse et al., DeepSeek, Model Distillation, and the Future of AI IP Protection, Fenwick (Feb. 3, 2025), https://www.fenwick.com/insights/publications/deepseek-model-distillation-and-the-future-of-ai-ip-protection?.
[39] INSAIT and LatticeFlow study finds EU AI regulatory flaws in DeepSeek, Inst. for Comput. Sci., Artificial Intelligence and Tech. (Feb. 4, 2025), https://insait.ai/deepseek-evaluation/.
[40] AI Act, European Comm’n, https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai (last visited Mar. 16, 2025).
[41] Id.
[42] INSAIT, supra note 39.
[43] Id.
[44] Regina S. Penti, et al., DeepSeek: Legal Considerations for Enterprise Users, Ropes & Gray (Jan. 29, 2025), https://www.ropesgray.com/en/insights/alerts/2025/01/deepseek-legal-considerations-for-enterprise-users.
[45] Id.
[46] Id.
[47] Id.