On Monday, Jan. 27, the release of DeepSeek R1, an open-source large language model (LLM) from China, caused around $1 trillion of stock market value in artificial intelligence (AI) companies to disappear in under 24 hours. AI chipmaker Nvidia lost almost $600 billion in market value alone.

DeepSeek produced their AI model at a fraction of the cost of American competitors like OpenAI — the creator of ChatGPT — and Google.  In a post on X, AI researcher Andrej Karpathy wrote that DeepSeek spent around $6 million on training their AI model — around a thousandth of the estimated $3 billion dollars OpenAI has spent on training ChatGPT. 

But according to the Financial Times, OpenAI claims to have evidence that DeepSeek cut financial corners by using distillation, which is a process by which DeepSeek uses OpenAI’s models to train their own. 

Even if DeepSeek did not use American models as a starting point, some argue that their cost numbers should not be taken at face value. “[DeepSeek does] not necessarily disclose all the investment that has gone into it,” Dr. Kian Beyzavi said in an interview with The Urban Legend. Beyzavi is vice president of AI strategy and operations at Danaher Corporation, an AI life sciences company.

According to Beyzavi, many of DeepSeek’s reported numbers likely refer to a single part of creating a model rather than the entire process. “It may be that to run inference and reasoning, [DeepSeek] used $5 million, but in fact, far more investment had gone into pre-training [and] accessing the data,” Beyzavi said.

Substantially lower costs compared to American models is not the only thing that stood out to investors. DeepSeek R1 is an open-source AI model, meaning that its code is freely available for public use. Other programmers can iterate upon it with their own code. Programs that are open source are typically free because the code is not proprietary to a single company. A lack of revenue from the monetization of LLMs could mean that investors might not immediately return on their initial investments. 

While OpenAI offers a free model, their high-end model costs $200 per month. “I’m really happy that DeepSeek is open source because I think it’ll discourage [other companies’] payment models to a degree,” Brody Izuel ’27 said. 

Another notable difference compared to ChatGPT, which is trained on virtually all text on the internet, is that DeepSeek is trained on fewer parameters. Developers can train AI models like DeepSeek R1 to serve as aids in areas of study that use specific skill sets, such as chemistry or calculus. “You can train it on differential equations, [for example],” Beyzavi said. “There isn’t a universe of answers; there’s a finite number of answers. … What DeepSeek has shown is that they can solve very specific problems — cheaper, faster, not necessarily better, but on par with ChatGPT.” This raises questions about the dominance and viability of American AI companies’ expensive generalized models.

American AI companies Oracle and OpenAI, as well as Japan’s Softbank, announced a new $500 billion venture called Stargate in a press conference with President Donald Trump, just days before DeepSeek R1’s release. The project is an effort to fund data centers, new electricity power sources and expensive chips used to train LLMs, called graphic processing units (GPUs). 

“This Stargate $500 billion [is] to say large language models aren’t going away and we need to keep expanding the training,” Beyzavi said. While DeepSeek R1 is a large language model, Stargate focuses on funding models with larger parameters and more intensive energy requirements.  

The use of AI in the geopolitical race between the U.S. and China has also come to the forefront. Beyzavi said, “If DeepSeek, in fact, ends up being better than OpenAI, what does that mean? What exactly has been lost?”

Some students might have concerns about government influence on LLMs. According to CNN, Deepseek displays characteristics of censorship that is prevalent in China. For example, according to The Guardian, DeepSeek is unable to mention the 1989 Tiananmen Square Massacre.

If Chinese AI models gain traction, large sources of information will be censored — regardless of whether the user is in China or elsewhere. Carl Haidamus ’25 said, “I think censorship could be problematic in the future because AI is advertised as the future. So if tools we use to access information are controlled by a government, that allows the government to push their agenda forward.”