Chinese artificial intelligence startup DeepSeek is developing its own custom inference chips to decrease its reliance on Nvidia and Huawei hardware. According to three people familiar with the matter, the company has ramped up its recruitment of chip-design engineers and is currently in discussions with foundry and memory partners to support its semiconductor ambitions.
Why is DeepSeek shifting to custom hardware?
DeepSeek’s move toward internal chip design aims to solve a critical bottleneck: the availability of high-performance hardware under strict U.S. export controls. By developing chips optimized specifically for inference—the stage where AI models generate responses—the company hopes to gain greater control over its infrastructure, according to sources cited by Reuters.
The company has historically relied on Nvidia’s H800, a chip specifically modified for the Chinese market, and more recently, Huawei’s Ascend processors. While Huawei’s chips were instrumental in the training of DeepSeek’s V4-Flash model, the startup’s pivot to in-house design follows a broader industry trend. Tech giants like OpenAI have recently moved toward custom hardware, such as the Jalapeno chip developed with Broadcom, to optimize performance and reduce dependence on general-purpose GPUs.
Inference chips are often cheaper and more energy-efficient than general-purpose GPUs because they are fine-tuned for specific, repetitive tasks rather than the intensive, broad-spectrum requirements of model training.
How does this impact the Chinese AI market?
DeepSeek’s expansion into hardware adds competitive pressure to an already crowded domestic market. Huawei currently holds approximately 50% of the $50 billion Chinese AI chip market, according to industry estimates, but that dominance is facing challenges from other tech firms like Alibaba and Baidu, which are also developing proprietary silicon.
The transition is not without significant risk. Designing competitive AI chips requires years of capital-intensive development. Furthermore, U.S. export restrictions prevent Chinese firms from accessing the most advanced overseas foundries and high-bandwidth memory, both of which are essential components for high-end AI inference hardware.
What are the primary hurdles for DeepSeek?
The company faces two major structural barriers to success, according to industry reporting:
- Manufacturing Constraints: U.S. bans currently restrict Chinese access to the world’s most advanced semiconductor fabrication facilities.
- Resource Access: Curbs on high-bandwidth memory limit the ability of domestic designers to build chips that can keep pace with international standards.
Despite these challenges, DeepSeek has signaled a shift in its business model. After years of avoiding external investment, the company was slated to raise $7 billion in a funding round in June, valuing the firm between $52 billion and $59 billion, according to Reuters.
Frequently Asked Questions
Why is DeepSeek building its own chips?
DeepSeek is developing its own chips to reduce its dependence on Nvidia and Huawei hardware and to optimize performance for the “inference” stage of AI processing, where models generate user responses.
What is the difference between inference and training chips?
Training chips are designed for the heavy lifting of building an AI model from scratch, while inference chips are designed to be more power-efficient and cost-effective for running models that have already been trained.
Are U.S. sanctions affecting DeepSeek’s development?
Yes. U.S. export controls limit access to advanced Nvidia chips, high-bandwidth memory, and top-tier foreign foundries, forcing Chinese companies like DeepSeek to prioritize domestic alternatives.
Stay updated on the latest shifts in semiconductor strategy and AI infrastructure. Subscribe to our newsletter for weekly insights into the tech industry.