The DeepSeek model developed by Chinese AI startup, DeepSeek, utilizes NVIDIA's (NVDA, Financial) H800 graphics processing units (GPUs), which are priced at approximately 70% of the top-tier version. The H800 serves as a scaled-back variant of the H100 processor commonly used by American companies. According to industry insights, the H800's computing speed is roughly half of the H100.
Pricing information reveals that the H800 processors cost between RMB 190,000 to 240,000, equating to 70-90% of the H100's price, which is RMB 260,000. For the development of the DeepSeekV3 model, 2,000 to 3,000 H800 processors are utilized, resulting in a total GPU cost of RMB 385 million to 721 million. However, DeepSeekV3's declared development expense of $5.576 million does not include GPU costs.
The model is designed to enable AI learning across approximately 2.8 million machine hours, with a cost of $2 per hour, merely one-tenth of the expense associated with U.S. models. It is suggested that leasing GPUs might help DeepSeek reduce these costs further.
