Facing US export controls, China's DeepSeek plans to make its own chips
Summary
Chinese large language model startup DeepSeek is reportedly planning to enter the silicon business, focusing on data center chips for inference. This strategic move, which has been in development for about a year with meetings with potential partners and engineer hirings, aims to reduce DeepSeek's reliance on both Huawei and Nvidia. US export controls have significantly limited Nvidia's presence in China, where Huawei currently controls approximately half of the data center chip market. DeepSeek's initiative mirrors similar efforts by US-based AI companies; for instance, OpenAI recently announced "Jalapeño," its first inference chip developed with Broadcom, and Anthropic is also exploring custom chip designs. These companies seek Apple-like control over their tech stack and an advantage in a constrained data center compute market.
Key takeaway
For Directors of AI/ML managing large-scale model deployments, you should actively assess your organization's reliance on external chip suppliers. The trend of major AI labs developing custom inference silicon, driven by export controls and compute scarcity, signals a critical shift. Consider initiating internal R&D into custom chip solutions or diversifying your hardware partnerships to secure future compute capacity and maintain competitive advantage.
Key insights
AI companies, facing supply constraints and seeking control, are increasingly developing custom inference chips.
Principles
- Geopolitical factors drive AI hardware strategy.
- Vertical integration enhances control and resilience.
- Custom silicon optimizes AI inference at scale.
Method
The article describes companies hiring engineers and meeting partners to develop custom data center chips for inference, aiming to reduce reliance on external suppliers.
In practice
- Evaluate custom chip development for inference.
- Diversify AI hardware supply chains.
- Investigate vertical integration for AI stack.
Topics
- AI Chips
- Data Center Inference
- US Export Controls
- DeepSeek
- OpenAI Jalapeño
- Vertical Integration
Best for: CTO, VP of Engineering/Data, Entrepreneur, AI Hardware Engineer, Director of AI/ML, Investor
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by AI - Ars Technica.