A Chinese start-up startup is grabbing the giant global tech firms by the horns, which has surprised the worldwide markets. DeepSeek has disrupted the foundations of major American companies such as Nvidia, OpenAI, Google, and Meta, all of whom have dominated the AI sector for years. Microsoft’s stock dropped 3.8%, TSMC’s U.S. stock slid 14%, and shares of AI chip pioneer Nvidia fell 16%.
DeepSeek appears to have the advantage of being a low-cost model and, most crucially, open source. However, the reality of DeepSeek’s claim about low cost is already being debated in the AI community.
Elon Musk, CEO of Tesla and X, has suggested that DeepSeek, an upstart artificial intelligence business, may have approximately 50,000 NVIDIA Hopper GPUs. This contrasts with DeepSeek’s claim to have 10,000 A100s.
Musk’s perspective was expressed in response to previous comments by Alexandr Wang, the millionaire CEO of Scale AI, who also indicated that DeepSeek most certainly has 50,000 NVIDIA H100 GPUs.
DeepSeek claims it took only two months and cost less than $6 million to create an AI model utilizing Nvidia’s less powerful H800 CPUs.
“Did DeepSeek build OpenAI for $5 million? Of course not,” Bernstein analyst Stacy Rasgon said in an interview to CNBC. “They did not train these reasoning model for 5 million dollars, we don’t know what their spending is to teach their reasoning model,” said Rasgon.
Rasgon also said that they are good models but the type of things they did were not miracles, all labs are working on these and deploying them, so nothing here was unknown to anybody.
DeepSeek has introduced a new generation of AI models, focusing on efficient and cost-effective hardware development, directly competing with established US companies.
The 2023-founded start-up has disputed the notion that scaling AI necessitates significant investment and processing capacity by claiming that its AI models either meet or surpass leading U.S. competitors at a tenth of the cost.
Subho Moulik, Founder and CEO, Appreciate is also of the view that DeepSeek’s claim of building its model for a fraction of the cost – $6 million compared to over $100 million to train the latest version of ChatGPT – is yet to be verified.
DeepSeek’s first significant feature was the introduction of new API features on July 25, 2024. Since then, they have released DeepSeek-V2.5, DeepSeek-R1-Lite, DeepSeek-V2.5, DeepSeek-V3 and DeepSeek APP.
The world took notice of DeepSeek when the DeepSeek-R1 version was published on January 20, 2025. Many consider January 20, 2025 as the Sputnik Moment for the AI industry after the release of the DeepSeek-R1 version.