In the competitive and vibra scene of artificial ielligence, we have always witnessed the emergence and fall of new actors. As one of these emerging actors, the Chinese company Deepsik has been focused on the focus of ambitious claims of dramatically reducing the cost of training for advanced artificial ielligence models. The company claims that it has trained its R1 with a $ 5 million budget using 4 GPUs. While American competitors are incurred by multiple dollars and over billions of dollars to achieve similar models. But how much are these claims reliable?
According to a report released by the Semianalysis Research Institute, Dipsic invests about $ 1.5 billion to create its own infrastructure and uses a powerful set of 6,000 Hopper GPUs, including 6,000 H800 and 6,000 H100. These data provided by Semianalysis coradicts Dip -Sick’s initial claims of low costs of model training. It seems that the $ 5 million figure only part of the cost of education specifically reflects the cost of graphic processing at the pre -training phase, and has ignored the heavier research and developme costs, mass processing of data and extensive infrastructure.
It is worth noting that Deepsic is actually coming to the heart of the Chinese High-Flyer Chinese investme company. The company had made significa investmes in the areas of artificial ielligence and graphics processors for many years. In year 2, High-Flyer decided to establish Dipsic as an independe and specialist in the field of artificial ielligence. One of the distinct features of Dipsic is the adoption of a self -approach to many similar startups. Instead of dependence on cloud services, Deepsik guides its dedicated data ceers. This independence enables dipsic to fully manage the testing and optimization process of its artificial ielligence models and to make the necessary changes quickly without the need to ieract with external companies.
Another key strength of Deepsic is the ability to attract elites and superior tales from the China borders. The company is offering promine artificial ielligence specialists from prestigious Chinese universities, such as Beijing University and the University of Zajiang by offering competitive and tempting compensatory packages. Reports show that some Dipsic artificial ielligence researchers receive $ 1.5 million annually, even beyond the paymes of large Chinese artificial ielligence companies such as Monshat.
Instead of focusing on increasing the hardware scale, Dipsic implemes a differe strategy and emphasizes the promotion of algorithms and technical innovations. For example, the company has developed the Multi-Head Late Atteion algorithm (MLA), which has been the result of mohs of research effort and widespread use of graphics processors. The CEO of Deepsik pois out that by adopting clever approaches and efficie algorithms, one can achieve more or even superior results with more limited resources.
However, Dip -Sick’s achievemes are not only the cause of technical innovation, but also the massive investmes and the attraction of elite tales also play a decisive role. From the perspective of many hardware experts, the company’s initial claims about the low cost of teaching artificial ielligence models are more promotional and far from reality. In fact, Like other leading artificial ielligence companies, Deepsik requires huge investmes and persiste efforts to maiain their competitive position.
25,199,000
23,899,000
Toman
Source: Tom’s Hardware
