DeepSeek’s launch of its new open-source AI reasoning model, R1, triggered a sell-off in Nvidia’s stock and propelled its consumer app to the top of the app stores.
Last month, DeepSeek announced that it had successfully trained a model utilizing approximately 2,000 Nvidia H800 GPUs over a period of nearly two months, incurring a cost of around $5.5 million. Recently, the company released a paper demonstrating that the performance of its latest model is on par with the most sophisticated reasoning models available globally. These advanced models are being developed in data centers that are investing billions in Nvidia’s high-speed, expensive AI chips.
The response from the technology sector to DeepSeek’s high-performance, cost-effective model has been overwhelmingly positive. For example, Pat Gelsinger expressed his enthusiasm on X, stating, “Thank you DeepSeek team.” Gelsinger, the former CEO of Intel, is a hardware engineer and currently serves as the chairman of Gloo, a startup focused on messaging and engagement for churches, which is preparing for an IPO. He departed from Intel in December after four years, during which he attempted to compete with Nvidia using Intel’s alternative AI GPUs, the Gaudi 3 AI.
Gelsinger emphasized that DeepSeek’s achievements should remind the tech industry of three crucial lessons: reduced costs lead to broader adoption; creativity thrives under limitations; and “Open Wins.” He asserted that DeepSeek will play a pivotal role in transforming the increasingly insular landscape of foundational AI model development, contrasting it with the closed-source approaches of OpenAI and Anthropic.
Gelsinger informed TechCrunch that R1 has made such a significant impact that Gloo has opted not to pursue or invest in OpenAI. Instead, Gloo is in the process of developing an AI service named Kallm, which will include a chatbot and various other functionalities.
“My Gloo engineers are currently utilizing R1,” he stated. “They could have utilized o1, but their access to o1 is limited to the APIs.” In approximately two weeks, Gloo anticipates having completely reconstructed Kallm “with our own foundational model that is entirely open source,” he added. “This is quite exciting.”
He expressed his belief that DeepSeek will render AI so affordable that it will not only be ubiquitous but also of high quality. “I desire enhanced AI in my Oura Ring. I want improved AI in my hearing aid. I seek more AI in my smartphone. I want superior AI in my embedded devices, such as the voice recognition in my electric vehicle,” he remarked.
Gelsinger’s enthusiastic response appeared to contrast with the sentiments of others who were less optimistic about the emergence of a higher-performing and more cost-effective competitor to reasoning foundational models. The trend in AI has been towards increasing costs rather than decreasing them. Some critics suggested that DeepSeek must have manipulated its figures in some way, implying that training expenses must have been higher. Others speculated that it could not claim the use of advanced chips due to U.S. export restrictions on AI chips to China. Additionally, some analysts pointed out areas where other models outperformed DeepSeek. There are also those who believe that OpenAI’s forthcoming model, o3, will significantly surpass R1 upon its release, restoring the previous balance.
Gelsinger dismisses these concerns. “Complete transparency will never be achievable, especially since much of the work was conducted in China,” he noted. “Nevertheless, all indications suggest that their training costs are 10 to 50 times lower than those of o1.”
DeepSeek demonstrates that progress in AI can be achieved “through engineering ingenuity rather than merely increasing hardware capabilities and computational resources. This is indeed exhilarating,” he concluded.
