Jensen Huang, co -founder and CEO of Nvidia Corp. , During a press conference in Taipei, Taiwan, on Tuesday 4 June 2024. NVIDIA is still working on the process of issuing certificates for Samsung Electronics Co. Memory chips, a final step required before the Korean company can start providing a basic component to train artificial intelligence platforms.
Annabel Chih Bloomberg Gety pictures
Nafidia The “AI Advancement” R1 “AI AIDSEK” is called, although the Chinese start starting, causing the price of the chip maker’s share by 17 % on Monday.
“Deepseek is an excellent progress of artificial intelligence and an ideal example for the test time,” CNBC spokesman told CNBC on Monday. “Deepseek’s work shows how new models can be created using this technology, taking advantage of the extensively available models and an account that is fully controlling the export.”
The comments come after Deepseek released last week R1, an open source thinking model that outperforms the best models from American companies such as Openai. The self -reported training cost for R1 was less than 6 million dollars, a small part of the billions spent by Silicon Valley companies to build artificial intelligence models.
The NVIDIA statement indicates that he sees Deepseek’s completion that it creates more work for American Chip Maker graphics units, or graphics processing units.
“Inference requires large numbers of NVIDIA graphics and high -performance networks,” the spokesman added. “We now have three scaling laws: pre -training and post -training, which continue, and expand the new test time.”
Nafidia also said that Dibsic’s graphics processing units were fully compatible. The meters are Amnesty International Alexander Wang Comments on CNBC last week It is believed that Dibsic used the NVIDIA graphics processing models that were banned in China the mainland. Dibsic says it uses special versions of NVIDIA graphics units for the Chinese market.
Analysts are now asking whether capital investments are billions of dollars such as companies Microsoftand Google and Dead For NVIDIA’s infrastructure is wasted when the same results can be achieved at a cheap price.
Earlier this month, Microsoft said it spent 80 billion dollars on the infrastructure of Amnesty International in 2025 alone, while the CEO of Meta Mark Zuckerberg said last week that the social media company plans to invest between 60 to 65 billion dollars at expenditures Capitalism in 2025 as part of the artificial intelligence strategy.
“If stereotypes prove that they are significantly lower, we expect a cost -cost benefit in the near term for advertising and travel and other consumer applications that use artificial intelligence services Justin Post books, Bofa Securities analyst, in Monday note.
NVIDIA’s comment also reflects a new topic discussed by the CEO of Nvidia Jensen Huang, CEO of Openai Sam Altman and Microsoft Satya Nadella in recent months.
Many of the mutation of artificial intelligence and the demand for NVIDIA graphics processing units was driven by the “scaling law”, a A concept in developing artificial intelligence Openai researchers proposed it in 2020. This concept indicates that the best AI systems can be developed by expanding the scope of the account and data that entered into building a new model, which requires more and more chips.
Since November, Huang and Altman have focused on new wrinkles of the scaling law, which Huang calls “measuring test time”.
This concept says that if the trained artificial intelligence model is full training, it spends more time using additional computer energy when making predictions or creating a text or pictures to allow it to “mind”, then it will provide better answers than it will be if he works for a lesser period.
The forms of the timing time law are used in some Openaii models Like O1 As well as the Deepseek model model R1.
He watches: The fund manager says that the difficult feeling of challenging from the exceptional markets
