Jensen Huang, co -founder and CEO of NVIDIA Corp., was held at a press conference in Taipei, Taiwan in Taipei, Taiwan on Tuesday, June 4, 2024. NVIDIA is still committed to memory chips for Samsung Electronics’s high models. This is the last step before the South Korean company began to provide an indispensable component for training AI platforms.
Annabelle Chih | Bloomberg | Getty image
Nvidia Although the appearance of Chinese startups has led to the stock price of chip manufacturers on Monday, Deepseek’s R1 model “is an excellent AI progress”, although the emergence of Chinese startups has led to 17 % of the stock price of chip manufacturers.
NVIDIA spokesman said to CNBC on Monday: “DeepSeek is an excellent AI progress, and it is also the perfect example of testing time expansion.” “Deepseek’s work illustrates how to use this technology to create new models, and uses a wide available model that fully conforms to export control. And calculation. “
After DeepSeek released R1 last week, these comments were published. This is an open source reasoning model. According to reports, the model is better than the best model of American companies such as OpenAI. The training cost of R1 self -reporting is less than $ 6 million. This is a small part of Silicon Valley’s billions of dollars spent to establish artificial intelligence models.
NVIDIA’s statement shows that Deepseek’s breakthroughs have created more jobs for the graphic processing unit or GPU of American chip manufacturers.
The spokesman added: “Reasoning requires a lot of NVIDIA GPUs and high -performance networks.” “We now have three zoom laws: training before and after training and new testing time.”
NVIDIA also said that the GPU used by DeepSeek conforms to export. Alexandr Wang’s CEO AI CEO Last week’s comment on CNBC He believes that Deepseek uses the NVIDIA GPU model banned in mainland China. Deepseek said it uses a special version of NVIDIA’s GPU for the Chinese market.
Analysts are now asking whether it is investment billions of dollars from companies such as companies Microsoft,,,,, Google and Yuan For NVIDIA -based AI infrastructure, when the same results can be achieved cheaper, it will be wasted.
Earlier this month, Microsoft said yes Expenditure 80 billion US dollars In only AI infrastructure in 2025, Mark Zuckerberg, CEO of Meta, said last week that the social media company plan Investment of 60 to 65 billion US dollars As part of the AI ​​strategy, in the capital expenditure in 2025.
“If the cost of model training is proven to be significantly reduced, we expect advertising, travel and other consumer application companies that use Cloud AI services to obtain recent cost returns, and long -term super standard AI and AI -related revenue and costs may be Luo.
NVIDIA’s comment also reflects a new theme of NVIDIA CEO Jensen Huang, OPENAI CEO Sam Altman and Microsoft CEO Satya Nadella discussed in recent months.
Most AI prosperity and the demand for the NVIDIA GPU are driven by the “Expansion Law” The concept of artificial intelligence development The concept was proposed by OpenAI researchers that the concept shows that it can develop a better AI system by large expansion of the calculation and data of new models, and more and more chips are required.
Since November, Huang and Ultraman have been focusing on the new wrinkles of the law of zooming, and Huang is called “test time scaling”.
This concept says that if a comprehensive training AI model spends more time using additional computer power when making prediction or generating texts or images to allow them to allow them to “rational”, then it will provide less time than running than running. Answer.
Some models of Openai use the form of test time expansion law For example O1 And Deepseek’s breakthrough R1 model.
watch: Fund manager said
