Recent claims about the Chinese startup DeepSeek using only $6 million to build an AI model within two months shocked developers. However, new facts are emerging about the actual cost of training such a model.
The Real Expenses of DeepSeek's AI Training
David Sacks, White House AI and Crypto Czar, challenged the assertions concerning DeepSeek's costs. According to a report by semiconductor analyst Dylan Patel, the training expenses for this model are actually over $1 billion. The $6 million figure only covers the final training stage, excluding capital and R&D expenditures.
SemiAnalysis Cost Breakdown
SemiAnalysis claims the $6 million refers only to the initial GPU training phase. Total infrastructure, R&D, and server capital expenses are estimated at $1.3 billion. Notably, DeepSeek employs various GPU models, including H100, H800, and H20. The expert MLA technology reduces costs by decreasing cache usage by 93.3%.
Impact on the Crypto Industry
Experts suggest DeepSeek's innovations could significantly impact the crypto sector, prompting a demand for increased efficiency and cost-effectiveness. This might stimulate new technologies and scaling approaches across various crypto projects. AI token shares have already shown a rise following recent reports.
Regardless of DeepSeek's model training costs, its innovative approach could unlock new possibilities for the crypto industry, demanding more focus on efficiency and expenses from developers.
