DeepSeek Clarifies True Cost of Training AI Model R1

DeepSeek has clarified the true cost of training its AI model R1, following confusion over initial reports. The company initially stated that the training cost was $294,000, but this figure only accounted for the reinforcement learning phase. The complete training process, including the foundational work on the V3 base model, actually cost around $5.87 million.

The misunderstanding arose from supplementary information released with a paper in January, which detailed the use of 512 Nvidia H800 GPUs for 198 hours to train the preliminary R1-Zero release. However, this did not include the extensive pre-training of the V3 model, which required 2.79 million GPU hours.

DeepSeek's R1 model was trained using a technique called Group Relative Policy Optimization (GRPO), a form of reinforcement learning. This method was applied after the base model was already developed, which involved significant computational resources and costs. The company has emphasized that the $294,000 figure was only for the reinforcement learning phase, not the entire training process.