What is DeepSeek AI and how is it disrupting the AI sector?

Nvidia stock plunged 17% on Monday, January 27, 2025, the steepest drop since March 2020, taking $589 billion off the company's market capitalization. What triggered the greatest single-day market cap fall in the history of the US stock market? The answer is DeepSeek, a Chinese AI firm whose latest model, DeepSeek R1, has not only matched but sometimes outperformed the performance of industry giants such as OpenAI and Google while operating on a shoestring budget.

While Nvidia has long been the darling of the AI revolution due to the extreme need for GPUs, Deepseek AI’s Deepseek R1 and Deepseek V2 innovations have proven that the future of AI might not rely on expensive hardware but on smarter, leaner algorithms. DeepSeek is 20 to 50 times cheaper to use than the OpenAI O1 model, depending on the task, according to a post on DeepSeek’s official WeChat account.

The USA-China rivalry adds another layer of intrigue, as Deepseek’s rise is seen by some as a symbolic victory for China’s tech ambitions, leaving American giants scrambling to keep up. It’s a classic David vs. Goliath story, except Goliath just got a multi-billion-dollar paper cut.

What is DeepSeek AI?

DeepSeek AI is a cutting-edge machine learning model built by Liang Wenfeng, a Zhejiang University graduate and co-founder of the quantitative hedge fund High-Flyer, focusing on natural language processing (NLP), data analysis, and decision-making. Unlike its Western peers, DeepSeek has a unique training style that prioritizes efficiency and cost-effectiveness.

• DeepSeek's models, including DeepSeek-V3 and DeepSeek R1, have an MoE architecture that activates just the processing blocks required for a given task, minimizing energy consumption and operational expenses.
• DeepSeek R1 improves its reasoning abilities through reinforcement learning, allowing it to find and enhance approaches over time.

DeepSeek's models are open-source, making them more accessible to a wider audience, and they cost a fraction of what competitors charge. How low are we talking about? The training cost for DeepSeek-V3 was apparently under $6 million, which is extremely inexpensive when contrasted to the hundreds of millions spent by competitors such as OpenAI on GPT-4, which costs more than $5.4 billion per year.

DeepSeek LLMs

• DeepSeek Coder: Released in November 2023, this is the first open-source model to be developed by DeepSeek for coding applications. It offers a highly efficient code generator, debugger, and review tool for the developers.
• DeepSeek LLM: Released in December 2023, the first general-purpose model developed by the company, providing a wide variety of natural language processing capabilities.
• DeepSeek-V2: Released in May 2024, DeepSeek-V2 focuses on high performance and reduced training costs. It is an architecture of Mixture-of-Experts (MoE) with a total of 236 billion parameters and 21 billion parameters turned on per token. Context length of 128,000 tokens is achieved, and it is successfully applied in text generation, coding, and conversation.
• DeepSeek-Coder-V2: Released in July 2024, this model is a 236 billion-parameter model, as a solution to encoding challenges. The context window size is 128,000 tokens, which is capable of dealing with big codebases.
• DeepSeek-V3: DeepSeek-V3 emerged last December, and it is a 671 billion-parameter model with a context length of up to 128,000 tokens. It is based on an MoE architecture so that it can perform various tasks efficiently. It is pre-trained on 14.8 T tokens and thus possesses wide coverage in many domains of knowledge.
• DeepSeek-R1: Released in January 2025, DeepSeek-R1 extends the work of DeepSeek-V3 and is targeted at complex reasoning tasks. It matches or surpasses the performance of leading models like OpenAI's o1 while maintaining a significantly lower cost structure. It also drives the chatbot in DeepSeek's, which has directly challenged ChatGPT.
• Janus-Pro-7B: Launched in January 2025, Janus-Pro-7B is a vision model that can interpret and generate images. It broadens DeepSeek's capabilities beyond text-based jobs.

OpenAI (ChatGPT) vs. DeepSeek

Controversy: The Solid Hype vs. What's Real!

OpenAI may have been the first to capture the world's imagination with its ChatGPT series, but DeepSeek has proved that the one who shuts the door doesn't necessarily shut out the best. The DeepSeek-V3 model, in particular, gives a punch to those hardware-heavy approaches adopted by OpenAI.

If we go by existing off-the-shelf frameworks depicting highly inflated analysts' views on OpenAI, it should always be intensified cargo-delivering robust concerns on a tech rival. Very recently, OpenAI faced a severe banter—the shift from the open-source ethos to one overly driven for corporate profits, between corporate hubris—the other model, DeepSeek, born into turmoil due to an alleged case of overhyping and utilization of High-Flyer connections to draw investments.

Are DeepSeek’s models truly revolutionary, or is this just another case of smoke and mirrors? It is reported that the AI chatbot tends to avoid answering the questions that project China in a bad light; for example, when asked about the Tiananmen Square Massacre, the chatbot initially generated an answer before abruptly replacing it with an error message that read, "Sorry, that's beyond my current scope. Let's talk about something else." Or a question like, is Arunachal Pradesh a part of India? Gives the same error. Indicating bias in the LLM.

The truth is, we don’t know yet. But what we do know is that DeepSeek is forcing OpenAI to up its game—and that’s a win for the entire AI sector.

Privacy Concerns Surrounding DeepSeek AI

The rapid expansion of DeepSeek AI, a Chinese AI business, has raised serious privacy concerns in the United States, notably over its data collection techniques and the ramifications of storing user data in China. Here's an in-depth look at the issues:

Data Gathering and Storage

• Voluminous Data Collection: DeepSeek accumulates lots of personal data, including email addresses, telephone numbers, date of birth, history of conversations, audio input, and even typing patterns. The information is stored on "secure servers" in China, creating privacy and security concerns.
• Lack of Transparency: The privacy policy is unclear on data retention lengths, anonymization procedures, and whether user data is used to train AI models. This opacity has raised concerns regarding the retention and use of personal information.
• Data Sharing: DeepSeek distributes user data with third parties, such as advertising and analytics partners, without giving opt-in or opt-out alternatives, further limiting users' control over their personal information.

Security and Compliance

• Inadequate Security Features: DeepSeek's security measures seem insufficient, with no specific mention of encryption of data being transmitted or stored, and there is no mention of two-factor authentication or any other sophisticated protections against unauthorized entry.
• Non-Compliance with International Privacy Laws: DeepSeek is not compliant with international privacy regulations such as the General Data Protection Regulation (GDPR), resulting in criticism from European consumer groups and the Italian Data Protection Authority.
• Data Breach: A recent data breach in which DeepSeek exposed one million sensitive records has raised alarms regarding the safety of user information.

National Security Implications

• Government Access: In China, firms are legally required to aid and collaborate with state intelligence services, which implies that whenever the Chinese government requires access to their country's companies' user data, they will provide it. This differs from standard U.S. statutes, which often require federal authorities to get court orders or warrants before gaining access to data maintained by American technology corporations.
• U.S. Government Response: The US Navy has issued a warning to its members not to use DeepSeek's AI services for professional or personal use due to potential security and ethical problems. Furthermore, the White House has stated that it is assessing DeepSeek's implications for national security.

User Privacy Risks

• Behavioral Surveillance: DeepSeek's collection of keystroke patterns or rhythms, similar to biometric data, may uniquely identify people, raising concerns about the extent of behavioral surveillance.
• Data Usage for AI Training: User inputs are included into datasets to further train AI models, which may expose personal information to unexpected purposes.
• Cybersecurity Risks: DeepSeek has revealed that it was targeted by a large-scale breach that resulted in the release of over 1 million log lines and secret keys, forcing it to suspend new user sign-ups on its web chatbot interface and revealing faults in its cybersecurity architecture.

Public and Expert Opinions

• Privacy Experts: Experts such as Lauren Hendry Parsons from ExpressVPN and Dr. Richard Whittle from the University of Salford have expressed serious concerns about DeepSeek's privacy policy, specifically the possibility of government access to user data.
• Tech Industry Figures: While others, such as Perplexity CEO Aravind Srinivas, have attempted to assuage concerns by pointing out that DeepSeek's AI can be downloaded and run locally, overall concerns about data storage in China remain.

Final Thoughts

If you sit back and watch this David-Goliath drama, you will realize that DeepSeek AI is actually a global disrupter that breaks the whole artificial intelligence industry, and not just a Chinese business. However you look at it, whether as an inspired underdog or a rising menace, they have come to change the rules of the game forever.

While the dust hasn't yet settled for DeepSeek AI, Alibaba has launched its own artificial intelligence platform, [Tongyi Qianwen (Qwen2.5-Max)](https://www.alibabacloud.com/en/solutions/generative-ai/qwen?_p_lc=1 which they claim outcompetes DeepSeek V3 in benchmarks like Arena Hard, LiveBench, LiveCodeBench, and GPQA-Diamond. However, that discussion is for another blog.