TOP 10 – AI APPLICATIONS

REVENUE: US $ 200M EMPLOYEES: 200

CEO: LIANG WENFENG FOUNDED: 2023

Wenfeng’ s AI app DeepSeek has rewritten assumptions about what it costs to build frontier models. DeepSeek-V3 trained on 14.8 trillion tokens for US $ 6m, using 2.664m H800 GPU hours.

Compare that to the estimated US $ 100m OpenAI spent training GPT-4, or the compute Meta burned through for Llama 3.1 – DeepSeek used one-tenth the power. Perhaps most significantly, DeepSeek-V3 is available under the permissive MIT License.

If a well-funded Chinese hedge fund can achieve comparable performance at a fraction of the cost, the competitive advantage shifts from who has the biggest GPU cluster to who has the cleverest architecture. Liang Wenfeng, CEO, DeepSeek

WATCH NOW

What is DeepSeek? AI model basics explained

48 January 2025