DeepSeek, a Chinese artificial intelligence company, became the hottest topic in tech this week, with industry experts and Wall Street fixated on one key figure: $6 million.
In its latest research paper, DeepSeek revealed that training its newest AI model cost approximately $5.576 million, based on Nvidia GPU rental prices. However, the company clarified that this figure covered only the “official training” phase and excluded expenses related to prior research, architectural experiments, and data optimization.
Earlier this week, DeepSeek’s AI Assistant surged to the top of Apple’s App Store as the most-downloaded free app in the U.S., surpassing OpenAI’s ChatGPT. Meanwhile, global tech stocks took a hit, with chip giants Nvidia and Broadcom collectively shedding $800 billion in market capitalization on Monday.
A new report from SemiAnalysis, a semiconductor research firm, offered additional context on DeepSeek’s spending. The report estimated the company’s historical hardware investments exceeded $500 million, with significant R&D and operational costs. It also highlighted that training the model required a substantial amount of computing power to generate synthetic data.
By comparison, Anthropic’s Claude 3.5 Sonnet reportedly cost “tens of millions” to train, yet Anthropic secured billions in funding from Amazon and Google—underscoring the vast financial resources required to develop and sustain AI models. “AI companies must experiment, refine architectures, clean datasets, pay employees, and much more,” SemiAnalysis noted.
DeepSeek’s research paper does not provide an estimate of its total computing costs, and the company has yet to respond to requests for comment.
SemiAnalysis emphasized DeepSeek’s achievement, stating, “DeepSeek is unique in reaching this level of capability at such a cost. R1 is a very strong model, and catching up to the reasoning edge this quickly is objectively impressive.”
Industry analysts lauded DeepSeek’s rapid advancements, especially given the U.S. has imposed three rounds of chip export restrictions on China in as many years. These developments have raised concerns that the U.S. may be falling behind in the AI race—a market projected to surpass $1 trillion in revenue within the next decade.
Bernstein analysts reflected on the intense discussions surrounding DeepSeek, noting that reactions ranged from “That’s really interesting” to “This could be the death knell of the AI infrastructure complex as we know it.”
Founded in 2023 by Liang Wenfeng, co-founder of High-Flyer—a quantitative hedge fund specializing in AI—DeepSeek emerged from the fund’s research unit to focus on large language models and artificial general intelligence (AGI). The startup remains fully owned and funded by High-Flyer, according to analysts at Jefferies.
DeepSeek’s momentum accelerated earlier this month with the release of R1, its reasoning-based AI model, which directly competes with OpenAI’s models. Unlike many proprietary systems, R1 is open-source, allowing AI developers worldwide to build on it.
How to sign up for DeepSeek AI—beginning with ChatGPT’s competitor
However, like other Chinese AI models, DeepSeek’s chatbot imposes content restrictions on politically sensitive topics. Reports indicate that when queried about certain policies of Chinese leader Xi Jinping, the chatbot redirects users away from such discussions.
OpenAI CEO Sam Altman has publicly praised DeepSeek’s model but has also suggested that OpenAI has evidence DeepSeek may have improperly used its data during development.
Speaking at an OpenAI event in Washington, D.C., on Thursday, Altman acknowledged DeepSeek’s capabilities, stating, “It’s clearly a great model.” He added that the company’s success underscores the intensity of AI competition and the importance of ensuring democratic AI leadership. “This highlights the growing interest in reasoning-based AI and open-source innovation,” he said.