According to individuals familiar with the developments, Ant Group Co., a company backed by Jack Ma, utilized Chinese-manufactured semiconductors to develop methods for training AI models, potentially reducing costs by 20%.
Ant Group employed domestic chips, sourced from affiliates such as Alibaba Group Holding Ltd. and Huawei Technologies Co., to train models using a method known as the Mixture of Experts in machine learning. The outcomes were reportedly comparable to those achieved with chips from Nvidia Corp., specifically the H800, though this information remains unpublished.
Based in Hangzhou, Ant Group continues to use Nvidia for AI development but now predominantly relies on alternatives, including chips from Advanced Micro Devices Inc. and Chinese producers, for its newer models. This move places Ant in a competitive arena with Chinese and U.S. companies, which has intensified since DeepSeek demonstrated how efficient models can be developed for a fraction of the investment by firms like OpenAI and Google’s parent, Alphabet Inc. It highlights China’s effort to find local substitutes for advanced Nvidia semiconductors, with the H800 being a relatively powerful chip currently restricted by the U.S. from Chinese markets.
Ant Group published a research paper this month claiming that its models sometimes surpassed Meta Platforms Inc. in specific benchmarks, though Bloomberg News has not independently confirmed this. If accurate, these models could significantly advance Chinese AI development by reducing the cost of AI services and inferencing.
Amid substantial investments in AI, Mixture of Experts models have gained popularity. Entities like Google and DeepSeek, a Hangzhou startup, endorse this technique, which divides tasks into smaller data sets similar to having specialists handle each segment efficiently. Ant did not comment officially despite inquiries.
The training of Mixture of Experts models typically relies on high-performance chips like those from Nvidia, posing a financial burden that has hindered broad adoption by smaller firms. Ant Group, aiming to overcome this limitation, has sought more efficient methods to train large language models, indicating its objective to scale a model without high-end GPUs.
This endeavor contrasts with Nvidia’s strategy. CEO Jensen Huang maintains that demand for computing power will continue to rise, requiring larger, more complex GPUs, despite emerging efficient models like DeepSeek’s R1. Huang has emphasized the need for bigger, more capable GPUs to drive revenue growth.
Ant Group reported incurring approximately 6.35 million yuan ($880,000) to train 1 trillion tokens using advanced hardware, but its optimized approach reportedly reduces the cost to 5.1 million yuan with less sophisticated hardware. Tokens are essential information units that models use to learn and respond effectively to user queries.
The company aims to apply recent breakthroughs in its large language models, Ling-Plus and Ling-Lite, for industrial AI applications, notably in healthcare and finance. This year, Ant acquired the Chinese online platform Haodf.com to enhance its AI services in healthcare. Ant has also developed AI Doctor Assistant to aid Haodf’s 290,000 doctors with tasks like medical record management.
Additionally, Ant provides an AI “life assistant” app named Zhixiaobao, and a financial advisory AI service called Maxiaocai. Regarding English language comprehension, Ant’s paper suggests that the Ling-Lite model outperformed a Meta Llama model in a key benchmark. On Chinese-language benchmarks, both Ling-Lite and Ling-Plus surpassed DeepSeek’s models.
Robin Yu, the chief technology officer of Beijing-based AI solution firm Shengshang Tech Co., commented on the importance of real-world applications, emphasizing that overcoming a single challenge can symbolize substantial advancement.
Ant has made its Ling models open source, with Ling-Lite containing 16.8 billion parameters and Ling-Plus holding 290 billion parameters. These parameters function like adjustable settings that guide the model’s performance. In comparison, experts estimate that ChatGPT’s GPT-4.5 involves 1.8 trillion parameters. DeepSeek-R1 features 671 billion parameters.
Ant faced certain hurdles during training, particularly with stability. The paper noted that even minor changes in hardware or the model’s architecture could lead to issues such as increased error rates.
On Monday, Ant announced the development of large model machines focused on healthcare, already in use by seven hospitals and healthcare providers in cities like Beijing and Shanghai. These large models incorporate DeepSeek R1, Alibaba’s Qwen, and Ant’s proprietary models to facilitate medical consultancy.
Additionally, Ant has introduced two medical AI agents—Angel, which has served over 1,000 medical facilities, and Yibaoer, supporting medical insurance services. Last September, it launched the AI Healthcare Manager within Alipay, its payments app.
This article was initially published on Fortune.com.
