TL;DR
- DeepSeek uses Huawei’s Ascend 910C chip for inference while training its AI model on Nvidia’s H100 GPUs, reducing reliance on U.S. hardware.
- Huawei’s Ascend 910C is optimized for inference but lacks the power for training large models, making it a cost-effective alternative for deployment.
- Huawei is developing the Ascend 920C to compete with Nvidia’s upcoming Blackwell B200, signaling China’s push for AI chip independence.
- China’s AI sector is moving toward self-reliance, driven by U.S. sanctions restricting access to Nvidia’s latest AI chips.
DeepSeek, a Chinese AI startup, has gained attention with its large language model (LLM), DeepSeek R1. The model is recognized for its efficiency and cost-effectiveness compared to competitors.
A key detail is that DeepSeek relies on Huawei’s Ascend 910C chip for inference, while it was initially trained on Nvidia’s H100 GPUs.
This decision highlights China’s push for greater independence in AI hardware.
Nvidia for Training, Huawei for Inference
DeepSeek’s R1 LLM was trained on Nvidia’s H100 GPU, but for inference—the process of generating responses—it relies on Huawei’s Ascend 910C chip.
According to a report by @Dorialexander, this setup highlights DeepSeek’s ability to adapt and innovate despite U.S. sanctions that limit access to cutting-edge technology.
Also Read: ChatGPT Search is now free for all users
The Ascend 910C chip, while not the fastest available, is a key component in DeepSeek’s operations. It handles inference tasks efficiently, though it may face limitations with more intensive workloads.
This strategy offers several advantages:
- Lower costs: Nvidia’s H100 GPUs are expensive and in short supply. Using Huawei’s Ascend 910C for inference reduces reliance on costly hardware.
- Sanctions and supply chain resilience: U.S. restrictions have made it difficult for Chinese companies to acquire Nvidia’s latest chips. By turning to Huawei, DeepSeek aligns with China’s goal of self-sufficiency in AI hardware.
- Inference optimization: The Ascend 910C is designed for inference workloads, though it lacks the performance required for training large models.
Huawei’s Next Move: The Ascend 920C
Huawei is developing the Ascend 920C, which aims to rival Nvidia’s Blackwell B200. The Blackwell B200 is Nvidia’s upcoming AI chip, expected to power next-generation AI models.
If the 920C delivers on its promises, it could strengthen China’s position in AI hardware. While Nvidia remains the industry leader, Huawei is making progress, providing an alternative for companies facing restrictions on U.S. semiconductor technology.
The Growing AI Rivalry Between China and the U.S.
DeepSeek’s decision to use Huawei’s Ascend 910C reflects a broader shift: China is working toward self-reliance in AI infrastructure.
The U.S. has imposed export restrictions on advanced chips, limiting China’s access to Nvidia’s best AI GPUs. In response, Chinese companies are accelerating domestic semiconductor development.
Key outcomes of DeepSeek’s hardware choice:
- More Chinese AI firms may adopt domestic chips, reducing dependence on U.S. suppliers.
- Huawei’s AI chip division could expand, driven by demand from China’s AI industry.
- AI hardware competition is intensifying, with Huawei emerging as China’s primary alternative to Nvidia.
Takeaways
the upcoming Huawei 920C chip could further enhance DeepSeek’s capabilities. If the 920C delivers on its promise to match Nvidia’s Blackwell B200, it could eliminate current bottlenecks and position DeepSeek as a stronger competitor in the global AI market.
In the meantime, DeepSeek’s success with the Ascend 910C chip is a clear example of how necessity drives innovation.
Despite the challenges, the company has managed to create an AI model that’s not only effective but also cost-efficient, proving that even smaller players can make a big impact.
