Overview
- Reuters and other outlets reported Tuesday that DeepSeek has begun designing a proprietary chip aimed at inference, the stage when trained models generate user responses.
- The project is described as early-stage with the company quietly hiring chip-design engineers and holding talks with chip-design firms, foundries and memory suppliers.
- DeepSeek has used Nvidia GPUs and Huawei Ascend processors for prior models and sees an in-house inference chip as a way to reduce reliance on those suppliers.
- U.S. export controls on advanced foundries and high-bandwidth memory create major manufacturing and performance hurdles that make initial production likely to target the domestic Chinese market.
- The effort comes as DeepSeek opens to outside capital in a roughly $7 billion funding round and has already triggered a modest market reaction with Nvidia shares dipping about 2 percent.