While Nvidia rebounded in pre-trading on Tuesday, analysts noted a change in the perception of the role of the company in the expensive development of moderate AI by graphic processing units (GPU), threatening one of the most precious technological titans in the world.
What did Deepseek achieve?
Deepseek claims to have pre-formulated his V3 model on only 2,048 GPU Nvidia H800 over a period of two months, each chip costing about $ 2 an hour to work. The total cost of training for the V3 model was $ 5.5 million, with 2.8 million hours of GPU, much less than rival models.
Meanwhile, its open source model of reasoning, R1, published earlier this month, has demonstrated capacities comparable to those of the more advanced models of Openai, Anthropic and Google, but with significantly lower training costs.
Does Deepseek prove that Nvidia chips are not essential?
Not yet. In an interview in 2023 with the Chinese media position, the founder of Deepseek, Liang Wenfeng, said that the company had gradually accumulated a stock of more than 10,000 GPU Nvidia, making it one of the main owners IT resources among AI Chinese start-ups.