Ant Group Achieves 20% Cost Reduction in AI Training Using Domestic Chips

Ant Group Co., backed by Jack Ma, has achieved a significant milestone in artificial intelligence (AI) development by leveraging domestically produced semiconductors. The company has developed techniques for training AI models that could potentially reduce costs by 20%. This advancement is part of a broader initiative by Chinese companies to create local alternatives to advanced semiconductors, particularly those produced by Nvidia Corp., which are currently restricted from being sold in China.

Ant Group has been utilizing Chinese-made chips, including those from its affiliate Alibaba Group Holding Ltd. and Huawei Technologies Co., to train AI models using the Mixture of Experts (MoE) machine learning approach. This method involves dividing tasks into smaller sets of data, similar to having a team of specialists focusing on different segments of a job, thereby making the process more efficient. The results obtained from these domestic chips are reportedly comparable to those achieved with Nvidia's H800 chips, which are among the most powerful processors available.

Despite the progress, Ant Group continues to use Nvidia for some of its AI development needs but is increasingly relying on alternatives from Advanced Micro Devices Inc. and other Chinese chip manufacturers for its latest models. This shift underscores the company's commitment to reducing its dependence on foreign technology and exploring more cost-effective solutions.

The development of these AI models marks Ant Group's entry into a competitive race between Chinese and US companies in the AI sector. This race has intensified since DeepSeek demonstrated the capability to train models at a significantly lower cost compared to the billions invested by companies like OpenAI and Alphabet Inc.'s Google. Ant Group's research paper, published this month, claims that its models sometimes outperform those of Meta Platforms Inc. in certain benchmarks. If these claims are verified, Ant Group's platforms could represent a significant advancement in Chinese AI development, potentially reducing the cost of AI services and inferencing.

The MoE models, which have gained recognition for their use by companies like Google and the Hangzhou-based startup DeepSeek, typically require high-performing chips like Nvidia's graphics processing units (GPUs). The high cost of these chips has been a barrier for many smaller firms, limiting broader adoption. Ant Group has been working on methods to train large language models (LLMs) more efficiently, aiming to eliminate this constraint. The company's research paper explicitly states its goal to scale a model "without premium GPUs," highlighting its focus on cost-effective solutions.

Ant Group's breakthrough in AI development using Chinese-made semiconductors is a testament to the company's innovation and its strategic efforts to reduce reliance on foreign technology. This development not only positions Ant Group as a key player in the global AI race but also underscores the growing capabilities of Chinese companies in the technology sector.