Chinese startup DeepSeek is shaking up the AI industry with its latest models, claiming they rival or outperform leading U.S. models at a fraction of the cost.
Global AI circles took notice when DeepSeek revealed in a recent paper that training its DeepSeek-V3 model required less than $6 million in computing power using Nvidia H800 chips. This breakthrough has propelled DeepSeek’s AI Assistant, powered by DeepSeek-V3, to surpass ChatGPT as the top-rated free app on Apple’s U.S. App Store.
This success has sparked questions about the massive AI investments made by U.S. tech giants. It has also impacted the stock performance of major players, including Nvidia, as industry leaders reevaluate their strategies.
Why DeepSeek is Making Waves in the AI Industry
The launch of OpenAI’s ChatGPT in late 2022 triggered a race among Chinese tech firms to develop their own AI-powered chatbots. However, the release of China’s first ChatGPT equivalent by search engine giant Baidu left many disappointed, highlighting a gap in AI capabilities between U.S. and Chinese companies.
DeepSeek has completely shifted this narrative. Its models, DeepSeek-V3 and DeepSeek-R1, have earned praise from Silicon Valley executives and U.S. tech engineers. The company claims these models rival the most advanced offerings from OpenAI and Meta, while being significantly more cost-efficient. For instance, the recently launched DeepSeek-R1 is reportedly 20 to 50 times cheaper than OpenAI’s models, depending on the task, as stated on DeepSeek’s official WeChat account.
In an interview with CNBC, Scale AI CEO Alexandr Wang alleged that DeepSeek owns 50,000 Nvidia H100 chips. He suggested the company might avoid disclosing this information due to Washington’s export controls on advanced AI chips for Chinese firms. DeepSeek has not responded to these claims.
Meanwhile, Bernstein analysts noted in a research report that the startup’s total training costs for the V3 model remain unclear. Although DeepSeek stated it spent $5.58 million on computing power, analysts believe the actual costs were much higher. Additionally, the training costs for the equally celebrated R1 model have yet to be disclosed.
Creators of DeepSeek
DeepSeek, based in Hangzhou, has Liang Wenfeng, co-founder of the quantitative hedge fund High-Flyer, as its controlling shareholder, according to Chinese corporate records.
In March 2023, High-Flyer announced on its official WeChat account a shift in focus. It stated plans to move beyond trading and concentrate resources on building a “new and independent research group to explore the essence of AGI” (Artificial General Intelligence). Later that year, DeepSeek was born.
OpenAI defines AGI as autonomous systems that outperform humans in most economically valuable tasks. While it remains unclear how much High-Flyer has invested in DeepSeek, the two share a close connection. High-Flyer operates in the same building as DeepSeek and holds patents for chip clusters used in training AI models, as shown in corporate filings.
In July 2022, High-Flyer’s AI unit revealed on WeChat that it operates a cluster of 10,000 A100 chips, further underscoring its deep involvement in AI development.
Also Read About
Perplexity Introduces Its Assistant for Android
Trump Announces $500 Billion AI Partnership with OpenAI, Oracle, and SoftBank