The Chinese AI company Deepseek reportedly delayed the launch of its newest AI model after a failed attempt to train it using chips from domestic tech giant Huawei.
According to the Financial Times, Chinese regulators encouraged Deepseek to switch from Nvidia's leading chips to Huawei's Ascend processors after the release of its R1 model in January. That plan ran into trouble. Deepseek faced persistent technical issues with the Ascend chips while training its R2 model, the FT reports. Even with Huawei engineers on-site, the team couldn't complete a successful training run.
These problems forced Deepseek to fall back on Nvidia chips for the compute-heavy training process. The delays pushed back the model's launch from May and gave competitors a head start. As a workaround, Deepseek now uses Nvidia hardware to train its models, but relies on Huawei's Ascend chips for less demanding inference tasks. Industry sources told the FT that Chinese chips still lag behind Nvidia in stability, connectivity, and software quality.
Deepseek V3.1 targets new Chinese hardware
Despite these setbacks, Deepseek has released an updated version of its V3 model. As The Register notes, the new V3.1 was trained using a special data type called UE8M0 FP8. In a WeChat post, Deepseek said this data type was designed for a next generation of domestically produced chips, which would be released soon.
This suggests that more powerful Chinese accelerators could be on the way. Huawei's current top chip, the Ascend 910C, doesn't natively support the FP8 data type. The move from the previously used E4M3 format seems less about efficiency and more about future hardware compatibility. V3.1 builds on an earlier V3 checkpoint, but adds a hybrid reasoning mode.
Ritwik Gupta, an AI researcher at the University of California, Berkeley, told the Financial Times that Huawei is likely experiencing some "growing pains" with its chips, but said it's only "a matter of time" before the company catches up.
