A report from the backschid; Nvidia’s $ 1.5 billion and 6,000 Nvidia chips to develop artificial intelligence
The world of artificial intelligence has always witnessed the emergence of startups. In the meantime, the Chinese company Deepsic attracted a lot of attention with its bold claims to dramatically reduce the costs of advanced artificial intelligence models. The company claimed to have trained its R1 with only $ 5 million and 4.3 GPUs; While its US competitors spend billions of dollars to train similar models. Are these claims true?
According to the Semianalysis Research Institute, Deepsic has spent about $ 1.5 billion to develop its infrastructure and uses 6,000 Hopper GPUs, including 6,000 H800 and 6,000 H100.
The statistics of the Semianalysis Research Institute are in conflict with the initial claims of Dip Sick on the very low costs of training its models. It seems that the $ 5 million figure only refers to some of the cost of education, namely the cost of graphic processing for the model’s pre -training phase, and the costs of research, development, data processing and general infrastructure have not considered.
Deepsic has actually come from a Chinese investment company called High-Flyer, which has made huge investments in artificial intelligence and GPU for many years. The company launched Depsic as an independent and specialized company in the field of artificial intelligence.
Unlike many other startups, Deepsik runs its dedicated datacenters instead of cloud services. This allows it to completely control the process of testing and optimizing its artificial intelligence models, without the need to coordinate with other companies.
Another main feature of Deepsic is the attraction of superior talents from inside China. By paying the tempting salaries and benefits, the company attracts prominent artificial intelligence professionals from prestigious universities such as Beijing University and the University of Zajiang. Some artificial intelligence researchers in Deepsic are said to receive more than $ 1.5 million annually, which is even higher than that of large Chinese artificial intelligence companies, such as Monshhat.
Related Articles
Instead of relying on increasing hardware scale, Dipsic focuses on improving algorithms and technical innovations. For example, the company has developed the Multi-Head Late Center (MLA) algorithm that needed months of research and great use of GPUs. The CEO of Deepsic states that using clever approaches and efficient algorithms can achieve similar or even better results with less resources.
However, the success of Deepsic depends not only on technical innovations, but also on macro investment and the attraction of top talent. From the point of view of hardware experts, the company’s initial claims of very low costs of teaching artificial intelligence models have been largely promoted and far from reality. In fact, Like other large -scale artificial intelligence companies, Deepsik requires enormous investments and continuous efforts to compete in this field.