Low-Cost AI Sparks Market Shock
Latest News: When DeepSeek unveiled its DeepSeek AI training cost as part of its low-cost AI systems in January, global investors rushed to sell tech stocks, fearing the models could challenge industry leaders like Nvidia.
Company Remains Low-Profile
Since the announcement, founder Liang Wenfeng and his company have stayed largely out of the spotlight, releasing only a few product updates.
Nature Paper Reveals Training Costs
A new Nature article co-authored by Liang disclosed that DeepSeek’s reasoning-focused R1 model was trained for $294,000 using 512 Nvidia H800 chips. This detail was absent from earlier reports.
Comparison with OpenAI
OpenAI CEO Sam Altman previously said foundational AI model training costs were “much more” than $100 million, highlighting the striking contrast with DeepSeek’s budget claims.
U.S. Scrutiny of Chip Access
Some U.S. companies and officials have questioned DeepSeek’s claims about its DeepSeek AI training cost. The H800 chips were designed for China after Washington banned exports of Nvidia’s powerful H100 and A100 chips in 2022.
Allegations of H100 Access
In June, U.S. officials told Reuters that DeepSeek may have obtained large volumes of restricted H100 chips. Nvidia, however, said the firm used lawfully acquired H800 processors.
Acknowledging Use of A100 Chips
In supplementary documents, DeepSeek admitted to owning A100 chips, which were used in the early stages of R1 development. The full training ran for 80 hours on H800 clusters.
China’s AI Talent Magnet
Reuters reported that DeepSeek has managed to attract top AI talent in China partly because it operates an A100-powered supercomputing cluster, a rare resource among domestic firms.
