The Chinese start-up DeepSeek rattled tech investors shortly after the release of an artificial intelligence model and chatbot that rivals OpenAI’s products.

Its chatbot assistant hit the top of Apple’s app store last week, surpassing ChatGPT at one point.

DeepSeek claims that its training costs only totaled about $5.6 million, while OpenAI said back in 2023 that it cost more than $100 million to train one of its models.

DeepSeek claimed it was able to develop its models with fewer chips from Nvidia, the leading AI chip supplier and one of the stock market’s top performers. As a result, AI-related stocks declined, causing the major stock indexes to slide earlier last week, while Nvidia lost $600 billion in market cap. By the end of January, Nvidia’s stock was down 11%. 

Despite the company’s promise, DeepSeek’s arrival has been met with controversy. OpenAI claims that DeepSeek used its proprietary models when developing their product. Australia and Taiwan have banned government workers from using any DeepSeek products and services due to security concerns, while Italy removed DeepSeek products from Apple and Google stores. 

In the U.S., Texas has also banned government workers from using DeepSeek, while the U.S. Navy and the Pentagon have reportedly banned members and staff from accessing the technology. Republican Senator Josh Hawley has introduced a bill that would bar all Americans from importing any AI technology from China. 

What is DeepSeek-R1? 

DeepSeek-R1 is the name of the company’s open-source language model, which is a reasoning model. This type of model more closely resembles the way that humans think compared to early iterations of ChatGPT, said Dominic Sellitto, clinical assistant professor of management science and systems at the University at Buffalo School of Management. 

To understand how that works in practice, consider “the strawberry problem.” If you asked a language model how many “r”s there are in the word strawberry, early versions of ChatGPT would have difficulty answering that question and might say there are only two “r”s. One theory for this is because there are technically only two “r” sounds in the word, Sellitto said. 

But reasoning models will look at each letter, determine if it’s an “r” and then count how many total “r”s there are, he explained. 

OpenAI has also developed its own reasoning models, and recently released one for free for the first time.

While DeepSeek’s chatbot provides the same capabilities as ChatGPT, it will censor questions that are considered politically controversial in China, said S. Shyam Sundar, director of Penn State’s Center for Socially Responsible Artificial Intelligence.

For example, it will refuse to answer questions about Tiananmen Square protests in 1989, when China’s military killed demonstrators. 

Does DeepSeek pose security risks? 

DeepSeek says personal information it collects from you is stored in servers based in China, according to the company’s privacy policy. 

It collects any information you voluntarily provide when you sign up for its services, such as your email address; internet- or network-related information about you, such as your IP address; and information from outside parties, such as advertisers. 

For those who are tech savvy, Sellitto said you have the option to download DeepSeek’s model and run it on your own hardware, disconnected from the company.

But many casual users are using the company’s chatbot. You need to be conscious about the data you provide to any organization, not just DeepSeek, Sundar said. 

“You run the risk of exposing some of your personal information if you ask it for, you know, tips on managing your mental health issues,” Sundar said. 

Sites in general share your information with other sites and services, which can make it easier for cyber criminals to scam you, Sundar pointed out. 

Why DeepSeek is disrupting the tech industry

AI companies spend a lot of money on computing power to train AI models, which requires graphics processing units from companies like Nvidia, Sellitto said. 

But if you don’t need as much computing power, like DeepSeek claims, that could lessen your reliance on the company’s chips, hence Nivdia’s declining share price. 

However, DeepSeek may be more reliant on GPUs than tech investors initially thought. While DeepSeek is touting it only spent a mere $5.6 million on training, the research firm SemiAnalysis says the company spent $1.6 billion on hardware costs. 

DeepSeek says it took only 2,000 GPUs to train its AI model, but SemiAnalysis says the company still has an inventory of 50,000 Nvidia Hopper GPUs, which is a type of GPU used for data centers.

People can also download DeepSeek’s models without paying a license fee, which Sellitto thinks will encourage more organizations to build AI tools. And if more people use DeepSeek’s open source model, they’ll still need some GPUs to train those tools, which would help maintain demand — even if major tech companies don’t need as many GPUs as they may have thought.