Ant Group Co., supported by Jack Ma, announced that it is creating new methods for training artificial intelligence models using semiconductors made in China, and these advancements are anticipated to lower costs by 20 percent.
Ant utilized local chips, including those from its affiliate Alibaba and Huawei Technologies, to train models using the Mixture of Experts machine learning technique.
While Ant continues to use Nvidia for AI development, it is increasingly turning to alternatives such as Advanced Micro Devices Inc. and Chinese semiconductors for its newer models.
These models signify Ant’s entry into a competitive arena between Chinese and US firms that has intensified since DeepSeek showcased the ability to train models at a fraction of the cost compared to the billions spent by OpenAI and Alphabet Inc.'s Google. This highlights how Chinese companies are aiming to leverage local substitutes for the most sophisticated Nvidia chips. Although not the most advanced, the H800 is a fairly powerful processor that is currently restricted from export to China by the US.
This month, the company released a research paper asserting that its models have at times surpassed those of Meta Platforms Inc. in specific benchmarks, according to reports.
If these claims hold true, Ant’s platforms could represent a significant advancement for Chinese AI development by considerably reducing the costs associated with inferencing or supporting AI services.
With significant investments pouring into AI, MoE models have become a favored choice, gaining attention for their usage by Google and the Hangzhou startup DeepSeek, among others. This approach divides tasks into smaller data sets, analogous to having a team of experts each concentrating on a specific part of a project, thereby enhancing efficiency.
Also Read: Lessons from the Rewarding Investment Journey of Warren Buffett
Nevertheless, the training of MoE models usually depends on high-performance chips, such as those offered by Nvidia. Until now, the expenses have been prohibitive for many smaller companies and have hindered broader adoption. Ant has been exploring methods to train LLMs more effectively and eliminate this limitation. The title of its paper clearly indicates this ambition, as the company aims to scale a model “without premium GPUs”.