Overview
- DeepSeek has not shared pre-release access to its V4 model with Nvidia or AMD and instead gave domestic suppliers including Huawei an early optimization window, according to Reuters sources.
- A senior U.S. administration official said the forthcoming model was trained on Nvidia’s Blackwell chips, which current export controls bar from shipment to China.
- The Blackwell systems were likely run in a cluster at DeepSeek’s Inner Mongolia data center, and the lab may strip technical markers that reveal U.S. hardware use, the official said.
- Officials also believe the training relied on distillation from outputs of leading U.S. models built by OpenAI, Google, Anthropic, and xAI.
- The controversy is intensifying debate in Washington over recent H200 licensing decisions and the broader approach to restricting advanced AI semiconductors to China, with Nvidia and AMD declining comment and DeepSeek not responding.