Jensen Huang, co -founder and director of Nvidia Corp., during a press conference in Taipei, Taiwan, Tuesday, June 4, 2024, Nvidia is still working on the certification process for the high scope of Samsung Electronics Co. Memory chips, the final necessary step before the Korean company may start supplying the component key to training AI platform.
Annabelle Chih | Bloomberg | Getty Images
Nvidia He called the Deepseek model R1 “Excellent AI progress”, despite the appearance of Chinese startups, which is why the price of a shares of chip manufacturers fell 17%on Monday.
“Deepseek is a great progress of AI and the perfect example of scalating the examination time,” NVIDIA spokesman Nvidia said on Monday. “Deepseek’s work illustrates how new models can be created using this technique, using widely available models and calculated that are fully in line with export control.”
Comments come after Deepseek released R1 last week, which is a model of opening open code that allegedly surpassed the best models of US companies like Openi. The cost of training from R1 was less than $ 6 million, which is a fraction of billions that the Silicon Valley companies spend to build their artificial intelligence models.
The NVIDIA statement indicates that Deepseek’s breakthrough is seen as a creation of more work for graphic units for the processing of a American chip or GPU producing manufacturer.
“The conclusion requires a significant number of NVIDIA GPU and the networking of high performance,” a spokesman added. “We now have three scales laws: pre-treatment and post-traning, which continues, and a new scanning of time.”
Nvidia also said that the GPUs who used the Deepseek fully in accordance with exports. Opposed to the Alexandr Wang’s Alexandr CEO Comments on CNBC last week That he believed that Deepseek was using Nvidia GPU -O’s models forbidden in continental China. Deepseek says he used special versions of GPU NVIDIA intended for the Chinese market.
Analysts are now wondering if more billions of dollars of capital investments than companies like Microsoft,, Google and Target For AI infrastructure based in NVIDIA, it is spent when the same results can be achieved cheaper.
Earlier this month, Microsoft said it was spending $ 80 billion On AI infrastructure only in 2025, while Meta Meta Mark Zuckerberg CEO said last week that the social media company planned invest between $ 60 and $ 65 billion in capital expenditures 2025 as part of its AI strategy.
“If the cost of training the model is shown significantly lower, we would expect a short-term advertising fee, travel and other consumer applications companies using Cloud AI services, while long-term revenues and costs associated with AI to AI are probably lower,” he wrote, “he wrote,” he wrote, “he wrote Analyst Bofa Securities Justin Post in note on Monday.
Nvidia’s comment also reflects a new topic that they have been discussing in recent months that NVIDIA CEO is Jensen Huang, Openi Executive Director Sam Altman and Microsoft CEO of Saty Nadella.
Much of Ai Bum -ai demand for Nvidia GPUs have encouraged “Scalating Law” and Concept in AI Development It was proposed by Openi researchers in 2020. This concept suggested that better AI systems can be developed largely by expanding the amount of calculation and data that has entered the construction of a new model, demanding more and more chips.
Since November, Huang and Altman have focused on new wrinkles for the Skalament Law, which Huang calls “scaling of time time.”
This concept says that if completely trained Ai model spends more time using the extra power of the computer when designing or generating a text or image to enable “reason”, it will provide better answers than it would be if it would take less time.
The forms of the Test Time Scalar Law are used in some of the Openi’s models such as o1 as well as Deepseek’s breakthrough R1 model.
WATCH: Deepseek Challenging Feeling US Exception to Markets, says Fund Manager
