Deepseek fired on the table this week, causing billions of losses in many companies, especially in Nvidia which has lost more than $ 600,000 million during the night in market capitalization, which reduced it from the first position that she led to the third and to be overcome by Apple Y Microsoft.
The main and the only reason for what some consider the end of AI bubbleis the cost of having developed this model. According to Depseek, the company only needs $ 6 million To create it.
The Openai model has had a cost of $ 100 million. The difference between the two is abyssal and how could it be otherwise, the United States, like other companies, doubt a lot about the veracity of this data.
What GPUs were used to create Depseek R1?
The cost of your training is linked to the equipment used. According to the company, Deepseek R1 was trained with the GPU H800A graphic from Nvidia was launched on the market in 2023 and which was sold without limits to China until the restrictions of the administration of Joe Biden in 2024.
Based on this information, the market wondered why pay more for the Nvidia H100 if we can get the same result with the 2023 H800?
If the H800 makes it possible to obtain the same result as the H100, GPU much more expensive than the H800, it makes no sense to continue to bet on the latest generation of GPUS by NVIDIA. It was, and not another, the main reason for the fall in the value of the actions that we commented at the start of this article.
The United States does not trust
The reason to doubt this figure is due to the proper functioning of Deepseek R1, very similar to that of Openai with GPT 01 Capable of thinking about your answers, that is to say looking for the right option to the question contrasting information with a greater number of sources in order to offer the right answer.
To give light to this question, according to Bloomberg, the government of Donald Trump opened an investigation to find out if Depseek used NVIDIA H-100 GPUs for training, GPUS that Nvidia is prohibited from selling to China precisely for 'preventing advancing as it will advance because Nvidia rapid as he did in the field of artificial intelligence.
The US government wants to determine whether the company has used the intermediary to buy the NVIDIA H100 GPU, thus avoiding trade policies in the United States. According to Howard Lutnick, from the United States Commerce Department:
It must end. If they will compete with us, which compete, but stop using our tools to compete with us.
One of the countries that is one of the suspects is Singapore, a country that represents 20% of NVIDIA sales. To prevent China from using intermediate countries, it has increased the number of countries with which Nvidia cannot exchange more than 40 years, but without including in the Singapore list. Indeed, many companies have offices in this country they use for products sent from China to the rest of the world.
