Training DeepSeek might not have been as cheap as we thought

WhenDeepSeek, a rising AI company, announced that it had trained its large language model for just $6 million, it raised eyebrows. While $6 million is no small figure, compared to industry giants like OpenAI and Google, it is next to nothing. This is because many other companies who are building AI models have spent billions. The training costs of DeepSeek seemed shockingly low.

However, new reports suggest that the $6 million figure was misleading—and the actual cost may be much higher. According to a recent report from SemiAnalysis, the $6 million number only accounts for GPU time during pre-training. This means it doesn’t include expenses involved in research and development. It also doesn’t account for costs for data processing and refinement, infrastructure costs, along with fine-tuning and optimization.

This is similar to how companies price their products. Companies need to take into account the bill of materials, but they also need to factor in costs like marketing, R&D, staff salaries, taxes, and more before arriving on the final price.

Another key detail in the SemiAnalysis report is that DeepSeek uses NVIDIA H100 Hopper GPUs. These are some of the most advanced (and expensive) AI chips available. These GPUs are in high demand and can cost tens of thousands of dollars each.

Taking everything into account, DeepSeek’s true AI training cost could be as high as $1.6 billion. This is a sum that is more in line with what other top AI companies are spending. While DeepSeek’s initial claim suggested a new wave of low-cost AI development, the reality is that cutting-edge AI still requires massive investments. However, there’s no denying that the efficiency of DeepSeek’s AI model and how it could upend the AI market as we know it.

The post Training DeepSeek might not have been as cheap as we thought appeared first on Phandroid.