While the Chinese startup’s success caused Nvidia’s stock price to plummet 17% on Monday, Nvidia has acknowledged DeepSeek’s recently released R1 model as a major breakthrough in artificial intelligence.

According to an Nvidia official, “DeepSeek is a significant AI breakthrough and a strong example of Test Time Scaling.” “Their advancements demonstrate how cutting-edge AI models can be built using this approach, utilizing widely available computing resources while remaining fully compliant with export regulations.”

Last week, DeepSeek released its R1 model, an open-source AI system with a reasoning focus. According to reports, R1 has outperformed top models from American companies such as OpenAI. Surprisingly, DeepSeek was able to create R1 with a training expenditure of less than $6 million, which is a small portion of the billions of dollars that major Silicon Valley companies are spending on AI development.

According to Nvidia’s position, the success of DeepSeek may increase demand for its GPUs, which are essential for AI inference procedures.

The spokesman went on to say, “Inference heavily relies on Nvidia GPUs along with high-performance networking,” “We now recognize three key scaling principles: pre-training, post-training, and the newly emerging test-time scaling.”

Nvidia also disputed assertions made by Scale AI CEO Alexandr Wang, stating that DeepSeek’s GPUs were legally allowed for use in China. Wang implied on CNBC last week that DeepSeek had utilized limited Nvidia GPUs; however, DeepSeek maintains that it used specially authorized models made for the Chinese market.

Analysts are questioning if the significant investments in AI infrastructure made by firms like Microsoft, Google, and Meta are required in light of DeepSeek’s cost-effective AI methodology. Microsoft recently revealed intentions to invest $80 billion in AI infrastructure by 2025, and Mark Zuckerberg, the CEO of Meta, said the business may spend up to $65 billion on AI capital expenditures the following year.

“Industries like advertising, travel, and consumer applications that rely on cloud-based AI services could see immediate financial benefits if training costs turn out to be significantly lower than expected,” said Justin Post, an analyst at BofA Securities. “However, this could also lead to a decline in long-term AI infrastructure revenues and costs for major cloud providers.”

Industry heavyweights including Microsoft CEO Satya Nadella, OpenAI CEO Sam Altman, and Nvidia CEO Jensen Huang have highlighted a wider trend that is reflected in Nvidia’s most recent remarks.

Previously, a notion known as the “scaling law,” which was first proposed by OpenAI researchers in 2020, was primarily responsible for the AI boom and, in turn, the demand for Nvidia GPUs. According to this theory, increased processing power and data input could boost AI performance, which would increase demand for GPUs.

However, Huang and Altman have been concentrating on a novel extension of this idea since late 2023, which they call “test-time scaling.”

According to this method, fully trained AI models can use additional processing resources during inference to improve their reasoning skills and, over time, the caliber of their responses. This approach is used in DeepSeek’s R1 and OpenAI’s more recent models, such as o1, suggesting a possible change in AI development approaches.