DeepSeek, an underdog Chinese startup with a large language model boasting powerful performance at a fraction of competitors’ steep training costs, knocked OpenAI’s ChatGPT from its top position in the Apple App Store — a development that on Monday spooked investors enough to send US technology stocks plummeting.
DeepSeek claims its V3 large language model cost just $5.6 million to train, a fraction of ChatGPT’s reported training costs of more than $100 million. With comparable performance to OpenAI’s o1 model, a 95% cost cut may be especially attractive to cash-strapped companies looking to leverage generative AI (GenAI).
The development sparked a pre-market selloff for major AI players, including Nvidia, Microsoft, and Meta. Investors sold off around $1 trillion in tech stocks in pre-market trading alone, with the S&P falling 2.3% and Nasdaq dropping by nearly 4% before the opening bell. Nvidia, the world’s leading supplier of AI chips, fell more than 11% in early trading. Chip designer Arm, Broadcom, and Micron Technology also suffered losses.
In a research note, Wedbush analyst Daniel Ives wrote: “Clearly tech stocks are under massive pressure led by Nvidia as Wall Street will view DeepSeek as a major perceived threat to US tech dominance and owning this AI revolution.”
Chirag Dekate, vice president and analyst at Gartner, thinks Wall Street may have overreacted to the DeepSeek news. In an interview with InformationWeek, Dekate says developments that reduce training costs will have an overall positive impact.
“It’s not just model innovation, it’s a system innovation,” Dekate says. “The DeepSeek innovations are real, and they matter … Lowering the cost structures is a net positive for the overall industry … DeepSeek enables a pathway to utilize resource more productively. Meta, Microsoft, Google, OpenAI, and other AI innovators can utilize those underlying capabilities even better. That will likely define the future of GenAI.”
Why is DeepSeek a Potential Disrupter?
Businesses can take advantage of massive cost savings on DeepSeek’s application programming interface (API) that boast costs of $.55 per million input tokens and $2.19 per million output tokens, a fraction of OpenAI’s API pricing of $15 per million input tokens and $60 per million output tokens.
But those savings come at a price — experts say widespread adoption of a Chinese-made model could pose significant security risks.
“From a security standpoint, you’re not going to want people putting data into servers that are hosted in China – same problem people had with TikTok,” says John Pettit, chief technology officer at IT consultancy Promevo. “You don’t know how data is being used and where it’s going to go. Even deploying it locally, you have to worry about supply chain injection.”
National security concerns in November prompted a bi-partisan US congressional group to sound the alarm on China’s progress in AI. The US-China Economic and Security Review Commission called for a government-funded effort to quickly develop artificial general intelligence (AGI) before China. AGI, which promises language models that match or better human intelligence, could be harnessed as a powerful weapon and give the country that first develops the technology a huge geopolitical advantage.
And DeepSeek CEO Liang Wenfeng stated in a recent interview that developing AGI is a top priority. “Our destination is AGI, which means we need to study new model structures to realize stronger model capability with limited resources,” Wenfeng told Chinese publication ChinaTalk in a November interview.
The US also alleges China backed hacking group Volt Typhoon’s efforts to disrupt US critical infrastructure. “China remains the most active and persistent cyber threat to US government, private-sector and critical infrastructure efforts,” according to a blog post from the Cybersecurity & Infrastructure Security Agency (CISA), who warned of continuing state-sponsored security threats.
Despite lower costs, Dekate says, enterprises will not likely rush into using DeepSeek widely because of potential legal liabilities. “Enterprises should always be careful about creating external facing products that are produced by open-source models,” Dekate says, noting that enterprise grade AI models offer more guardrails, security, and higher quality outputs. “There are going to be constraints [with open source models] that Gemini, OpenAI and other models do not have… you are going to get a more comprehensive answer on certain topics.”