Deepseek is called revolutionary because it was so cheap to develop. However, the cost price is higher than expected.
If there was an AI company in recent weeks that started with the attention, then that was deep key. The Chinese company developed the R1 language model, which is used to drive a chatgpt-like AI application. Striking, because Chinese companies have no longer had access to processors with American technology for a long time. These are deemed necessary for the AI development. That the Chinese company nevertheless managed to train a language model therefore gets a lot of attention.
That is not the only thing that Deepseek received attention. The price tag that hung on the AI application could also be called at least impressive. Companies such as Google, Meta and OpenAi spend billions to train their AI applications. Deepseek would have only spent 6 million dollars on the R1 model. There was doubt about that claim. Rightly so, it turns out: Semianalysis calculated that the 6 million dollars only pays part of the development.
Material, Staff and R&D
The 6 million dollars that Deepseek mentions as a cost only relates to the GPU costs for the preparation. Material and personnel costs, or even R&D costs, do not include Deepseek in the cost. So wrongly, because that money was also spent. Moreover, those costs would have risen considerably: Deepseek used Nvidia H100 GPUs to train the AI model. Just about every AI company wants those GPUs to train AI applications. The question of it is very high and the pricing corresponding.
Everything taken into account would have spent about $ 1.6 billion on the development of the AI system. Although Deepseeks R1 model is efficient and efficient, it is therefore not nearly as cheap as the company first made it believe. As a result, the AI system is less revolutionary than first thought.
The Chinese AI system is also not particularly revolutionary in other areas. For example, the language model holds on to Chinese censorship rules and refuses to answer questions that may be sensitive. Even in terms of safety, the AI system did not score too good points. Databases in which conversations of users were stored were freely available for a while via the internet.