Deepseek’s R1 model, which has aroused a great impact in the world of artificial intelligence, was defined by Nvidia as “an excellent artificial intelligence progress”. This statement of Nvidia came at a time when Deepseek caused sales waves and Nvidia lost $ 600 billion a day. On the other hand, OpenAI CEO Sam Altman said “a very impressive model”.

Praise from Nvidia and OpenAI to Deepseek

OpenAI CEO Altman said in a statement that they now need more information processing power and new generation models will be amazing, “Deepseek’s R1 model is quite impressive especially in terms of their price. Of course we will offer much better models and it is really motivating to have a new opponent on the field! We will make some announcements soon. ” He said.

“Deepseek is an excellent example of Time Scaling, a perfect artificial intelligence progress and test time scaling. Deepseek’s work shows how new models can be created by using this technique, and how to utilize from widely used models and information processing technologies that are completely suitable for export control. ”

The R1, which was released last week, attracts attention with its low cost. According to Deepseek’s statements, the training cost of the model remained below $ 6 million. This figure is quite modest compared to the artificial intelligence models developed by the giants in the Silicon Valley by spending billions of dollars.

The R1 model also draws attention with the fact that US -based companies, such as OpenAI, leave behind or match the best models in terms of performance. Nvidia sees this development as a positive development as it will increase the demand for graphic processing units (GPU). Sözcü, “Inference processes require a significant amount of Nvidia GPU and high -performance network infrastructure. We now have three scaling rules: Before the ongoing training, post -training and new test time scaling. ”

Test Time Scale or Test Time Scaling (TTS) is an approach to improve the scalability or performance of a model in artificial intelligence and machine learning models at the stage of inference, ie, while predicting new data. We see this approach in “Thinking” models such as OpenAI O1, O3 and Deepseek R1.

Huge investments are questioned

Deepseek’s high performance at low cost makes the investments of large technology companies. Microsoft announced that it plans to spend 80 billion dollars for its artificial intelligence infrastructure in 2025, while Meta CEO Mark Zuckerberg announced that they have allocated a budget of 60 to 65 billion dollars for the same year.