For two months, a model known as ‘Owl Alpha’ operated anonymously on OpenRouter, quietly processing massive volumes of traffic—peaking at roughly 559 billion tokens per day—before its true origin was revealed. It was not the product of a specialized AI lab, but of Meituan, a food delivery and logistics giant with over 770 million users. The release of LongCat-2.0, a 1.6 trillion-parameter Mixture-of-Experts (MoE) model trained entirely on a cluster of over 50,000 Huawei Ascend chips, marks a pivotal moment in the global AI arms race. It serves as a stark demonstration that frontier-level training is no longer the exclusive domain of Western tech giants or specialized research institutions; it is becoming a function of capital, data, and the sheer scale of domestic infrastructure.
The technical specifications of LongCat-2.0 are formidable: 1.6 trillion parameters, an MoE architecture with approximately 48 billion parameters activated per token, and a 1 million token context window. Meituan reports a score of 59.5 on SWE-bench Pro, a metric that positions the model as a high-performing agentic coding tool. However, this figure remains vendor-reported and has not been independently verified or placed on a public leaderboard. While Meituan’s internal documentation compares this to GPT-5.5, that comparison relies on data that was not independently re-run, necessitating a cautious interpretation of its true competitive standing against Western frontier models.
US export controls, intended to throttle the development of advanced AI in China, have instead acted as a catalyst for a ‘sovereign compute’ pivot. By restricting access to NVIDIA’s most advanced hardware, Washington has forced Chinese firms to optimize software stacks for domestic silicon, effectively eroding the long-term lock-in of the CUDA ecosystem. NVIDIA’s China-specific market share has plummeted from approximately 95% in 2023 to roughly 55% by 2025. While NVIDIA’s moat remains real, it is shrinking, as the ‘good enough’ threshold for domestic chips rises rapidly, threatening the company’s pricing power in the world’s second-largest AI market.
This transition is supported by massive capital deployment. Chinese cloud service providers are projected to increase capital expenditure by roughly 65% in 2025, with top internet firms investing over $70 billion in AI infrastructure. Huawei is at the center of this supply chain, with plans to scale the output of its Ascend 910C chips to approximately 600,000 units in 2026. As domestic firms like ByteDance shift nearly 60% of their semiconductor orders to Chinese suppliers, the reliance on foreign hardware is being systematically dismantled. The massive capital injection into domestic infrastructure is currently struggling to expand High Bandwidth Memory (HBM) supply, formalizing a state of persistent scarcity.
Meituan’s pivot from logistics to agentic commerce—using LongCat-2.0 to power in-app assistants for restaurant recommendations and complex booking tasks—highlights how AI capability is being integrated into the fabric of daily economic life. The democratization of frontier training means that the ability to build and deploy large-scale models is increasingly decoupled from access to Western silicon. As the gap between NVIDIA’s absolute hardware monopoly and the efficiency of domestic alternatives narrows, the global AI landscape is fracturing into distinct, competing compute spheres.
The uncertainty surrounding the long-term reliability of large-scale Ascend clusters compared to NVIDIA’s H100 or H200 systems remains a critical variable. Export controls have reached a point of diminishing returns; they have successfully restricted NVIDIA’s reach while simultaneously incentivizing the very domestic infrastructure that threatens to render those restrictions obsolete. Managing the trade-offs between hardware scarcity, the limitations of domestic HBM supply, and the necessity of software optimization for non-NVIDIA silicon now defines the operational reality for Chinese AI developers.