Meituan trains record AI model with domestic Chinese chips

What happened?

Meituan, the Chinese local services platform (with over 700 million annual active users), has announced that its new AI model LongCat-2.0, with 1.6 trillion parameters, was trained end-to-end using exclusively domestic Chinese chips. According to the company, it is the first model of this size to achieve such a feat, representing a direct response to semiconductor export restrictions imposed by the United States since October 2022. The model was trained on a massive computing cluster that, according to internal sources, used a combination of Huawei's Ascend chips and processors from startup Enflame, although Meituan has not officially confirmed the suppliers. The training took approximately three months and consumed an estimated budget of tens of millions of dollars, far below the costs of similar Western models.

Why is it important?

Meituan's announcement is significant for several reasons. First, it demonstrates that China can develop and train cutting-edge AI models without relying on advanced chips from companies like NVIDIA, whose exports to China are severely limited by US sanctions. While NVIDIA dominates the AI training market with its A100 and H100 GPUs (banned for China), and the H100 delivers up to 2000 TFLOPS in FP8, domestic Chinese chips like Huawei's Ascend 910B achieve around 320 TFLOPS in FP16, a considerable gap. However, Meituan managed to scale training to 1.6 trillion parameters, competing with models like GPT-4 (estimated at 1.7 trillion parameters) and Google's Gemini Ultra. This suggests that the capability gap between China and the West in AI may be closing faster than expected, especially in training efficiency and software optimization.

Moreover, Meituan's achievement has geopolitical implications. US restrictions, including the ban on exporting advanced chips to China and the inclusion of companies like Huawei on the entity list, aimed to curb China's technological advancement. However, this milestone shows that sanctions are driving domestic innovation, forcing Chinese companies to develop alternatives. According to a report by the Center for Strategic and International Studies (CSIS), China has increased its semiconductor investment by 40% annually since 2022, and companies like Meituan are proving that investment is paying off.

Historical context

Since the US imposed export controls in October 2022, Chinese companies have struggled to access high-performance chips like the NVIDIA A100 and H100. In response, the Chinese government has pushed for semiconductor self-sufficiency, supporting companies like Huawei (with its Ascend chip) and startups like Cambricon and Enflame. Meituan, primarily known for its food delivery and local services platform (with over 6 million delivery drivers), has been heavily investing in AI to optimize its logistics, recommendations, and fleet management. In 2023, the company launched its first AI model, LongCat-1.0, with 100 billion parameters, trained on NVIDIA chips. The leap to LongCat-2.0 with domestic chips represents a significant strategic shift.

Historically, AI development in China has heavily relied on foreign hardware. For example, Baidu, Alibaba, and Tencent have primarily used NVIDIA chips to train their models. However, sanctions have accelerated the search for alternatives. In 2024, Huawei launched the Ascend 910C, which offers performance comparable to the A100 in certain tasks, and companies like Biren Technology have developed specialized AI chips. Meituan has been one of the first to integrate these chips at scale, which could serve as a test case for other Chinese companies.

Market consequences

This achievement could have multiple impacts:

• Global competition: Meituan could offer AI services to third parties, competing with cloud providers like Alibaba Cloud or Tencent Cloud. If LongCat-2.0 demonstrates competitive performance in reasoning and language generation tasks, it could attract Chinese companies seeking alternatives to Western models. Additionally, Meituan could license its efficient training technology to other companies, creating a new revenue stream.
• Supply chain: Demand for domestic Chinese chips could increase, accelerating innovation in national semiconductors. Companies like Huawei, Cambricon, and Enflame could see a rise in orders, which in turn would improve their economies of scale and reduce costs. According to IDC data, the Chinese AI chip market is expected to grow at a compound annual rate of 35% until 2028, reaching $50 billion.
• Regulatory pressure: The US could tighten restrictions further or seek new ways to limit Chinese progress. For example, it could expand sanctions to companies using domestic Chinese chips or impose tariffs on products containing such chips. It could also pressure allies like Japan and the Netherlands to strengthen their own export controls. However, this approach could backfire, accelerating Chinese self-sufficiency.
• Impact on NVIDIA: NVIDIA could lose a significant portion of the Chinese market. In 2023, sales to China accounted for approximately 20% of NVIDIA's revenue, about $10 billion. If domestic Chinese chips prove viable for training large-scale models, demand for NVIDIA GPUs in China could drop sharply, affecting its future revenue.

What readers should know

It is important to note that although the training was done with domestic chips, specific details about the model's performance compared to its Western counterparts have not yet been disclosed. Meituan has published some internal benchmarks showing that LongCat-2.0 outperforms GPT-3.5 on mathematical reasoning and reading comprehension tasks, but it has not provided direct comparisons with GPT-4 or Gemini. Additionally, the scalability of this solution to other Chinese players will depend on the availability and efficiency of local chips. Currently, production of Huawei's Ascend chips is limited by sanctions restricting access to advanced lithography, which could affect the ability to scale production.

Meituan has not specified which exact chips it used, but speculation points to Huawei's Ascend 910B or Enflame's chips, a startup that has developed specialized AI processors with performance up to 256 TFLOPS in FP16. It is also mentioned that Meituan used parallelization techniques and software optimization to compensate for hardware limitations, such as unified memory and efficient inter-node communication. However, these details have not been independently confirmed.

"LongCat-2.0 demonstrates that AI self-sufficiency is not just a desire but a reality under construction," says an industry analyst. "While performance is not yet comparable to the best Western models, the mere fact of having trained a 1.6 trillion parameter model with domestic chips is an impressive technical milestone."

In summary, Meituan's announcement marks a turning point in the AI race between China and the West. Although questions remain about actual performance and scalability, the message is clear: China is determined to achieve AI self-sufficiency, and US sanctions may be accelerating that process rather than hindering it. The coming months will be crucial to see if other Chinese tech giants follow Meituan's lead and whether LongCat-2.0's performance is validated in real-world applications.