Ant Group (Alibaba) achieves AI breakthrough: More cost-effective models through innovative chip combination

Published on: March 24, 2025 / Updated on: March 24, 2025 – Author: Konrad Wolfenstein

Ant Group (Alibaba) achieves AI breakthrough: More cost-effective models through innovative chip combination – Image: Xpert.Digital

China's technological breakthrough: AI models become 20% more cost-efficient

Trade conflict is driving technological progress in AI chips

Ant Group, a subsidiary of the Chinese technology giant Alibaba, has announced a significant breakthrough in the development of AI models. Through an innovative combination of semiconductors manufactured in China and the US, the company has succeeded in training its AI models 20 percent more cost-effectively. This development could have far-reaching consequences for the global AI chip market and strengthen China's position in the technological arena. Particularly noteworthy is that Ant Group is relying on less powerful but efficient alternatives to the high-priced Nvidia chips without sacrificing significant quality. This comes against the backdrop of increasing trade tensions between the US and China and represents an important step towards greater technological independence.

Related to this:

Day 1 after the AI earthquake on the stock market: Nvidia crash, German tech stocks in distress – what does this mean for your investments?

Ant Group's innovative chip strategy

The solution developed by Ant Group is based on a clever mix of different semiconductor types. Instead of relying exclusively on the high-performance but expensive Nvidia H800 chips, the company is now increasingly using more cost-effective alternatives from AMD as well as from Chinese manufacturers such as Alibaba itself and Huawei. The development of this alternative chip strategy was significantly accelerated by US export restrictions on high-performance chips to China.

For its AI training models, Ant Group uses the “Mixture of Experts” (MoE) approach, a technique that enables more efficient training of large language models. This method involves breaking tasks down into smaller datasets for more efficient processing – “similar to using a team of experts, each focusing on a specific part of the problem, to improve overall efficiency.” This approach has already proven successful for other companies such as Google and the Chinese startup Deepseek.

Technical performance and comparison with Nvidia

Despite switching to supposedly weaker chips, Ant Group claims to be achieving results comparable to those of Nvidia hardware. According to the company, it has reduced training costs from 6.35 million yuan (approximately €810,000) to 5.1 million yuan for one trillion tokens. This significant cost reduction, while maintaining the same performance, could increase the pressure on Nvidia as the dominant AI chip provider.

Ant Group recently published a research paper claiming that its models can outperform models from other major technology companies like Meta in certain benchmarks. Specifically, in Chinese benchmarks, its own language models, Ling-Lite and Ling-Plus, are said to perform better than comparable models from Deepseek.

Economic impact and cost optimization

Cost optimization through alternative chip strategies is becoming increasingly important in the AI industry. While prices for end-customer services like ChatGPT remain relatively stable, costs in the B2B sector are continuously decreasing. Alibaba itself had already reduced the prices of various AI models several times last year, including an 85 percent price reduction for its vision language model Qwen-VL.

These cost reductions are part of a larger trend in the AI industry, driven by various factors:

The intense competition between different providers, both in China (Alibaba, Tencent, Baidu, JD, Huawei and Bytedance) and globally
The development of smaller, more cost-effective models that are becoming increasingly powerful
New methods such as model distillation, in which the knowledge of a large model is transferred to a smaller one
The cost pressure from numerous open-source or freely available models such as Meta's Llama family

The growing importance of cost-efficient AI models

The breakthrough announced by Ant Group comes at a time when the efficiency of AI models is increasingly coming into focus. In January, the Chinese startup Deepseek caused a stir when it presented an AI model that was trained with significantly fewer resources than previous models. This development even led to a dramatic drop in Nvidia's share price.

Nvidia CEO Jensen Huang argues, however, that the real need for computing power arises not during training, but rather during the generation of responses. Nevertheless, the Ant Group's approach demonstrates that significant cost savings are possible even in the field of AI training.

Related to this:

AI models explained simply: Understand the basics of AI, language models, and reasoning

Geopolitical significance and technological sovereignty

The development of Ant Group has a significant geopolitical dimension. Strict US export restrictions on high-performance chips have increasingly led Chinese companies to seek alternatives to American technology. US Commerce Secretary Gina Raimondo made it clear in December 2023 that state-of-the-art AI chips should not be supplied to China and that Chinese industry should not be enabled to manufacture them domestically.

In this context, Ant Group's chip strategy represents an important step toward greater technological independence. Robert Lea, senior analyst at Bloomberg Intelligence, noted that if confirmed, Ant Group's claims would demonstrate "that China is on track to become self-sufficient in AI as it shifts to lower-order, computationally efficient models to circumvent export controls on NVIDIA chips.".

Impact on global technology competition

China is making significant efforts to become a global leader not only in electric vehicles but also in AI. Ant Group's innovative approach could serve as a model for other Chinese companies operating in an increasingly complex geopolitical environment. The ability to flexibly access diverse technologies could prove to be a crucial competitive advantage.

Recent developments show that China is making significant progress in AI despite US restrictions. Similar to Ant Group, Deepseek, with its R1 model, has demonstrated that it's possible to achieve similar performance to OpenAI's Chat GPT using older hardware and a fraction of the budget.

Broader implications for the global AI market

The breakthrough announced by Ant Group could have far-reaching implications for the global AI market. A price war among AI providers has been raging in China for months. The trigger was the release of "V2," the predecessor to DeepSeek's current AI, "V3." In response, Alibaba and other technology companies like Baidu and Tencent slashed prices for their respective AI programs by up to 97 percent.

Price pressure and optimization in a global context

Price pressure in the AI sector is not limited to China. Western companies like Google have also begun optimizing their AI models and lowering prices. For example, in September 2024, Google reduced the prices for its Gemini 1.5 Pro model for input and output tokens by more than 50 percent.

Developments at Ant Group and other Chinese companies could further reinforce this global trend. Particularly interesting is the fact that the more cost-effective models are not necessarily less powerful. Google, for example, reports significant performance improvements in its optimized models, especially in mathematics (20 percent improvement), long context, and vision.

Another turning point in AI development from China?

The breakthrough announced by Ant Group could mark a significant turning point in AI development. The ability to train powerful AI models with more affordable chips could challenge Nvidia's dominance in the AI chip market and accelerate the democratization of AI technologies.

This development also demonstrates how companies can strengthen their position in a dynamic and challenging market environment through innovative approaches and strategic partnerships. Utilizing a diversified chip base could be one way to overcome the challenges of global supply chains while simultaneously reducing development costs.

It remains to be seen whether other companies will follow Ant Group's example and develop similar strategies. The ability to train AI models cost-effectively could be a crucial competitive advantage in an increasingly competitive market and ultimately enable access to AI technologies for a wider range of organizations and applications.

Related to this: