DeepSeek Reportedly Designing Custom AI Chip to Bypass US Curbs
DeepSeek is designing its own artificial-intelligence chip, according to a Reuters report published on Monday that cited people familiar with the matter. The move would shift China’s most prominent AI lab from utilizing other companies’ silicon to creating their own.
Key Points:
- Design vs. Working Chip: While not yet a functional chip, DeepSeek aims to develop its own AI hardware.
- Chinese Lab’s Motivation: Driven by talent and motive, the Chinese lab seeks to build an AI chip independently.
- Shift from Nvidia Hardware: DeepSeek has transitioned from relying on Nvidia hardware to tuning models for Huawei’s Ascend processors and other domestic Chinese silicon.
- Chip Fabrication: The custom chip will be fabricated by Semiconductor Manufacturing International Corporation (SMIC), China’s largest foundry.
- Challenges: SMIC faces constraints due to US and Dutch export controls, limiting their access to advanced chipmaking tools and processes.
- Inference Chip: The reported chip is designed for inference tasks, not training, aiming to handle user queries efficiently.
- Market Demand: With approximately 70% of AI compute demand expected from inference, a purpose-built chip could be lucrative.
- Precedent: As early as 2025, reports indicated DeepSeek’s interest in chip design and potential reliance on SMIC for domestic processors.
- Obstacles: Designing competitive AI accelerators is a lengthy process, and China’s foundries currently lag behind TSMC in terms of yields and performance.