Meituan Tests 1 Trillion Parameter AI Model on Domestic GPUs

Meituan is quietly testing a trillion-parameter AI model, a move that signals China's growing independence in the high-stakes race for computing power.

The food-delivery and services giant’s new model, reportedly trained on a domestically developed cluster of tens of thousands of GPUs, represents a significant milestone for China's technology sector as it navigates restricted access to foreign-made semiconductors. According to Chinese media reports, Meituan has begun sending out invitations for testing, making it the country's first publicly acknowledged trillion-parameter model trained entirely on homegrown hardware.

This development arrives as Chinese technology firms are reportedly being discouraged by Beijing from purchasing even approved, lower-spec GPUs from American companies like Nvidia, as detailed in recent comments by U.S. Commerce Secretary Howard Lutnick. "This move by Meituan, if successful, could provide a blueprint for other Chinese tech giants to achieve cutting-edge AI capabilities despite the challenging geopolitical landscape," an analyst familiar with the matter said.

While Meituan has not disclosed the specific domestic GPUs used, the scale of a "trillion-parameter" model places it in the upper echelon of global AI development. For comparison, Tencent's recently released Hy3 Preview model has 295 billion parameters. While parameter count is not the only measure of a model's capability, a trillion-parameter model suggests a significant investment in training resources, likely costing tens of millions of dollars and requiring a massive, highly-coordinated hardware infrastructure.

For investors, Meituan's project is a crucial indicator of its long-term strategy to de-risk its AI development from geopolitical supply chain shocks. By building on a domestic compute stack, the company could sidestep U.S. export controls that have complicated access to Nvidia's market-leading GPUs. This vertical integration, while expensive, could secure a critical long-term advantage in a market where computing power is paramount.

The Economics of Self-Reliance

The decision to train a frontier model on a domestic GPU cluster is as much about economics as it is about national strategy. As explained in a recent Nebius whitepaper, the total cost of training is determined by infrastructure efficiency, not just the hourly price of a GPU. A large, 3,000-GPU cluster can cost over $6,000 per hour, and any downtime or inefficiency directly inflates the final bill. By developing a tightly integrated stack of domestic hardware and software, Chinese firms like Meituan aim to lower their total cost of ownership and accelerate training times, creating a durable competitive advantage.

This effort stands in stark contrast to the global market, where Nvidia commands roughly 92 percent of the data center GPU market. The U.S. chipmaker's H200 and Blackwell-series GPUs are the gold standard, but Washington has blocked their sale to China. The resulting vacuum has spurred local rivals like Huawei and accelerated the push for self-sufficiency, even if the domestic chips are not yet on par with Nvidia's top-tier products. Meituan's initiative suggests that for large-scale players, the performance gap may be narrow enough to justify building on a local foundation.

A New Front in the AI Race

Meituan's move pushes it directly into the arena with other major Chinese AI players. Tencent's Hy3 model recently demonstrated dramatic performance gains in coding and agentic tasks, with its score on the SWE-bench Verified benchmark jumping from 53% to 74.4%. Alibaba and other firms are also heavily invested in their own large-scale models. The development of a trillion-parameter model could significantly enhance Meituan's core services, from more personalized recommendations and logistics optimization to entirely new AI-driven products, strengthening its market position.

The focus on a domestic GPU cluster is the key differentiator. It signals a potential decoupling of China's AI ambitions from its reliance on American technology. For Meituan (03690.HK), this represents a massive capital expenditure but also a strategic moat. If the company can prove that large-scale, competitive AI can be built on a non-Nvidia stack, it could boost investor confidence in the long-term viability of the broader Chinese tech sector in the face of ongoing trade restrictions.

This article is for informational purposes only and does not constitute investment advice.