Jack Ma-backed Ant Group, has developed cost-cutting AI training techniques using Chinese-made semiconductors, reducing costs by 20%, according to sources familiar with the matter. The company utilized domestic chips, including those from Alibaba Group and Huawei Technologies, to train AI models using the Mixture of Experts machine learning method. These results were comparable to those achieved with Nvidia’s H800 chips, the sources revealed, requesting anonymity due to the confidential nature of the information. While Ant Group continues using Nvidia for AI development, it increasingly relies on alternatives, including chips from Advanced Micro Devices Inc. and various Chinese manufacturers, for its latest models.
The models signify Ant Group’s entry into the growing competition between Chinese and U.S. companies. This race has intensified since DeepSeek showcased how highly capable models can be trained at a fraction of the cost compared to the billions spent by OpenAI and Google’s Alphabet Inc. This highlights the efforts of Chinese firms to rely on local alternatives to Nvidia’s most advanced semiconductors. Although not the most cutting-edge, the H800 is a powerful processor currently prohibited from being shipped to China by the U.S. government.
This month, the company released a research paper claiming that its models have sometimes outperformed Meta Platforms Inc. in specific benchmarks, although Bloomberg News has not independently verified this. If the claims hold, Ant’s platforms could significantly advance Chinese AI development by drastically reducing the cost of inferencing and supporting AI services.
As companies invest heavily in AI, MoE models have gained popularity, with recognition for their use by Google and the Hangzhou-based startup DeepSeek, among others. This technique breaks tasks into smaller data sets, like having a team of specialists focus on different project parts, thus improving efficiency. Ant declined to provide further comments in an email statement.
Training MoE models usually depend on high-performance chips, such as the graphics processing units (GPUs) Nvidia sells. However, the high cost of these chips has been a barrier for many smaller companies, limiting widespread adoption. Ant has been focusing on methods to train large language models (LLMs) more efficiently, aiming to overcome this limitation. The title of its research paper reflects this goal, with the company seeking to scale models “without premium GPUs.”
Ant Group’s paper emphasizes the growing innovation and rapid technological advancements in China’s AI industry. If the company’s claim is verified, it would signal that China is making significant strides toward achieving AI self-sufficiency, as the country increasingly adopts lower-cost, computationally efficient models to bypass export restrictions on Nvidia chips.
