DeepSeek Tried to Train Its AI on Huawei Chips — It Didn’t Go Well

While US companies are at the forefront of the AI race, China isn’t too far behind. Due to the long-standing tensions between both countries, China has been forced to adapt and come up with locally-grown solutions as alternatives to the West. This includes AI models. While many Chinese AI models are gaining traction, it looks like DeepSeek is facing some issues. The company was due to launch its next-gen AI model, but that has been delayed due to DeepSeek attempting to use Chinese chips.

Next-gen DeepSeek AI model delayed because of chips

A report from the Financial Times has revealed that the launch of DeepSeek R2 has been delayed. R2 is supposed to be the next DeepSeek AI model, but apparently, the company tried to use local chips, leading to its delay.

For a bit more context, the majority of AI development and deployment is powered by the use of NVIDIA’s tech. However, the trade war between the US and China has resulted in the US government limiting and restricting the type of tech that can be exported to China. This means that instead of getting the latest NVIDIA chips, China ends up with lower-specced versions. This is because NVIDIA has to meet certain requirements to export its tech to China.

However, the Chinese government has been pushing local companies to use local tech. In DeepSeek’s case, the company attempted to train its R2 model using Huawei’s chips. Unfortunately, this resulted in persistent technical issues during the training process. Ultimately, DeepSeek was forced to use NVIDIA’s chips for training and Huawei for inference.

The company was supposed to launch R2 in May, but obviously, they missed it.

China’s semiconductor problem

If anything, this highlights the pickle China finds itself in regarding semiconductor technology. The US and its allies have taken various steps to limit China’s access. This includes accessing the advanced EUV lithography machine needed to make semiconductors.

Huawei has attempted to advance its semiconductor ambitions. To everyone’s surprise, the company has done pretty well for itself. They’re not exactly closing the gap, but given the restrictions, it’s impressive. However, this issue with DeepSeek R2 shows that there is only so far Huawei can go, at least for now.