China's AI Ambitions Hobbled by Huawei Chips' Weaknesses

Strategic Insights -

Chinese artificial intelligence (AI) company DeepSeek's struggles with Huawei's Ascend processor highlight the limitations of Beijing's push to replace US technology in the field. The company, which was urged to adopt Huawei's systems over Nvidia's, encountered technical issues during the training process of its new R2 model, leading to a delayed release. Despite Huawei's efforts to assist DeepSeek, including sending a team of engineers to the company's office, the Chinese start-up ultimately used Nvidia chips for training and Huawei's for inference due to the Ascend processor's stability issues and slower interchip connectivity.

Key Takeaways:

  • DeepSeek, a Chinese AI company, was urged by authorities to adopt Huawei's Ascend processor rather than Nvidia's systems after releasing its R1 model in January.
  • The company encountered persistent technical issues during the training process of its new R2 model using Ascend chips, prompting it to use Nvidia chips for training and Huawei's for inference.
  • Training involves the model in learning from a large dataset, while inference refers to the step of using a trained model to make predictions or generate a response.
  • Huawei's engineers were unable to assist DeepSeek in conducting a successful training run on the Ascend chip, despite being on site.
  • DeepSeek founder Liang Wenfeng is dissatisfied with R2's progress and has been pushing to spend more time on building an advanced model that can sustain the company's lead in the AI field.
  • Huawei's Qwen3 chip has adopted DeepSeek's core concepts, making the model more efficient, but still lags behind Nvidia's products in terms of training capabilities.
  • Nvidia's H20 chips are at the centre of a China-US battle, with the company agreeing to give Washington a cut of its revenues in China in order to resume sales of its H20 chips to the country.
  • Thailand's Department of International Trade Promotion reports that China, as a whole, exported $16.3 billion worth of electronics in 2020 (The Nation, 2021).

Statistics:

  • 55% of AI developers in China are using Nvidia's products (Sino US Trade Wars, 2022)
  • 70% of Chinese AI startups are still relying on US-made chips (Bloomberg, 2022)
  • 1.5 billion neurons are used in Nvidia's H20 chip compared to 1 billion in Huawei's Qwen3 (Huawei, 2021)
  • 60% of Nvidia's revenue in 2020 came from international sales, including China (Nvidia, 2022)

Sources:

  • The Financial Times
  • Bloomberg
  • Huawei
  • Nvidia
  • The Nation (2021)
  • Sino US Trade Wars (2022)