Why DeepSeek Is Rattling Markets, Explained

In the ever-evolving landscape of artificial intelligence, a seismic shift occurred last week when DeepSeek, a rising star from Hangzhou, China, unveiled its groundbreaking AI model: DeepSeek-R1. With little warning, this unassuming yet formidable contender sent ripples of disruption through an industry long dominated by Western giants. 

DeepSeek’s emergence has been described by some as a “Sputnik moment” for the AI industry, highlighting the rapid advancements in AI capabilities outside of the United States, and it showed on the markets: Share prices for U.S. chip-making giant Nvidia were down17 percent and the NASDAQ closed Monday down 3 percent, though both rebounded Tuesday, with Nvidia regaining 8.8 percent and the NASDAQ up 2 percent. At a speech to House Republicans on Monday, President Trump said, “The release of DeepSeek AI from a Chinese company should be a wakeup call for our industries that we should be laser focused on competing to win.” 

The emergence of DeepSeek raises questions about the evolution of the technology and presents concerns for U.S. competitors and national security. Here is what we know now.

What is DeepSeek and how is it different from other AI platforms like ChatGPT?

DeepSeek is based in Hangzhou, China, and owned by Liang Feng, a hedge fund investor. His previous venture was High-Flyer, a quantitative trading firm. These firms’ natural evolution toward AI is driven by the need for better predictions, adaptability, and automation in an increasingly complex market environment. AI provides the tools to analyze massive data sets, learn from patterns, and execute trades efficiently. Liang’s firm announced in March 2023 on its official WeChat account that it was going beyond its trading business to concentrate resources on creating a “new and independent research group, to explore the essence of AGI” (artificial general intelligence). DeepSeek was created later that year. Liang is not known for giving frequent public interviews, but he has occasionally shared insights in high-profile outlets as DeepSeek’s impact has gained global attention. His comments often focus on AI’s future and China’s role in global tech, reflecting his strategy of positioning DeepSeek as a serious competitor to U.S. firms.

DeepSeek stands out for its combination of accessibility, practicality, and efficiency. The product achieves similar levels of performance to OpenAI’s ChatGPT or Google’s Gemini, and many consumers will not notice much of a difference across use of these services. DeepSeek does have some accessibility advantages and supports a broader range of languages. The AI Assistant app, comparable to ChatGPT or Gemini, emphasizes real-world practicality: It integrates deeply with mobile operating systems, offering seamless voice-to-text, language translation, and context-aware suggestions. This is appealing to everyday users, not just tech enthusiasts or professionals. Finally, DeepSeek’s price and affordability is the real difference and achievement. The R1 model is about 90 percent cheaper to use than most other models while achieving similar levels of performance with innovative training methods and resource optimization. Imagine you have a big box of Legos, but you only need a few to build a small house. Instead of using all the pieces and making a huge mess, you pick just the right ones to build your house quickly and neatly. Resource optimization is trying to use less power, less memory, and just the right amount of work to get the best results.

What did the company do technically that makes this so different and important?

At the heart of DeepSeek’s technical innovation is a mixture-of-experts (MoE) architecture, a model design that activates only the necessary computing resources for each task. Imagine a user asks an AI assistant powered by DeepSeek to translate a document from English to Mandarin. A traditional AI model, like OpenAI’s GPT-4, would activate its entire neural network for the task, processing every aspect of the translation using the same set of parameters, leading to high computational costs. With DeepSeek, the model intelligently selects only the necessary “expert” components of its network—experts in linguistics, grammar, context, and domain in this case. The system thinks for a shorter and more direct period of time, using less compute and fewer cycles of GPUs to figure out what the best possible answer is. This approach dramatically reduces energy consumption and computational costs while boosting efficiency.

What is ‘open-source’ and why did Deep Seek choose this model?

Open-source refers to software whose source code is made publicly available for anyone to inspect, modify, and distribute. This means developers around the world can collaborate, improve, and customize it. Popular examples include the Linux operating system and the Firefox web browser. In the context of AI, Meta’s product called Llama is open-source. Closed-source, or proprietary, software, is owned by a specific organization or individual. The source code is kept secret and only authorized parties can modify or distribute it. Examples of closed-source software include Microsoft Windows and Adobe Photoshop. In the context of AI, OpenAI’s ChatGPT product is closed-source. Open source has historically been better for safety, security, innovation, startups, and protection from sovereignty given the details are out in the open and not behind closed doors.

DeepSeek has embraced open-source accessibility, making its models freely available under one of the most permissive open-source licenses in the industry. This move should democratize AI development, allowing researchers and developers worldwide to build upon its technology—an approach starkly different from the closed, proprietary models favored by most of its competitors. Per a popular open-source community called HuggingFace, there are already 500 derivative models of DeepSeek, including those from well-known companies like Perplexity and Groq. Many are hosted in the U.S. and free to use. 

Should I download the DeepSeek AI Assistant app?

The DeepSeek AI Assistant jumped to the top of the Apple App Store charts this week, surpassing OpenAI as the top AI chat app. Per its own terms of use and privacy policy, the company tracks everything, including your inputs, your device metadata, your behavioral data, and beyond. This comes with serious security and privacy risks. All data goes back to servers in China. Tread lightly. David Sacks, the Trump administration’s new AI and crypto czar, has correctly laid out the risks and does not recommend downloading the app.

What are some of the national security and broader concerns as DeepSeek gains steam?

The major topic of discussion in many circles this week centered around the cost of training the DeepSeek model and more specifically how it relates to U.S. export controls. DeepSeek claims that it was able to train the R1 model for $6 million, whereas comparable models have cost north of $80 million. When reading the fine print, DeepSeek states that the $6 million figure “does not include costs associated with prior research and ablation experiments on architecture, algorithms, and data.” It seems unlikely that the number is simply $6 million, and we are left to believe that, similar to other AI companies, a significant amount of capital was expended prior to this moment to build out infrastructure. We have some clues—but also some questions. 

High-Flyer’s AI unit said on its official WeChat account in July 2022 that it owned and operated a cluster of 10,000 Nvidia A100 chips. By October 2022, the U.S. government had banned the export of high-performance AI chips, including Nvidia A100 and H100, to China to limit its AI and supercomputing capabilities. In October 2023, further restrictions were imposed, preventing Chinese companies from acquiring even downgraded versions of these chips (such as the A800 and H800, which were originally designed to comply with earlier restrictions). Scale AI CEO Alexandr Wang said during an interview with CNBC last Thursday that DeepSeek somehow now has 50,000 Nvidia H100 chips, which he claimed would not be disclosed because that would violate Washington’s export controls that ban such advanced AI chips from being sold to Chinese companies. Nvidia later disputed the claim. 

The ability to access advanced AI hardware, such as Nvidia’s chips, is a crucial factor in the development of state-of-the-art AI models. For companies like DeepSeek, these restrictions present a challenge to developing competitive technologies without relying on hardware from U.S. companies. While DeepSeek has found ways to leverage available hardware, questions remain about how effectively it complied with the latest export regulations. Beyond these regulations, Bloomberg reported Tuesday that Microsoft and OpenAI were also investigating whether data output from OpenAI’s technology was obtained in an unauthorized manner by a group linked to DeepSeek. This situation adds to the complex regulatory environment surrounding the rapidly evolving AI industry, where global competition and national security concerns are increasingly intertwined. 

DeepSeek has demonstrated that scaling up AI models relentlessly and building out massive computing infrastructure may not be the only path toward artificial general intelligence. The open-source nature of the product, the cost efficiency, and the novel structural breakthrough evolve the industry in new and creative ways. Its challenge to the dominance of Western companies in the AI space is very real this week and should have major impacts as the industry continues to mature.

How does this affect the new ‘Stargate’ announcement from President Trump?

On January 21, President Trump announced the launch of Stargate, a joint venture among OpenAI, SoftBank, Oracle, and the investment firm MGX. The initiative plans to invest up to $500 billion in artificial intelligence infrastructure in the United States over the next four years, with an initial commitment of $100 billion. The project aims to construct 20 data centers, each approximately 500,000 square feet, with the first facilities already under construction in Texas. These centers are expected to create over 100,000 jobs and bolster the nation’s AI capabilities.

The DeepSeek R1 model demonstrated that high-performance AI can be achieved with lower costs and resources, potentially rendering large-scale infrastructure investments less critical. It is unlikely the spending numbers, the jobs created, and the square footage of those investments will change. However, there will likely be more scrutiny of how the dollars will be spent. What DeepSeek has shown is that you can get a lot more out of each chip than expected. You can get more AI. This is wonderful as we start to use AI for more things and should be a net benefit to the capital investment Stargate is making.