How Meituan Trained a 1.6-Trillion Parameter AI Model Without Nvidia Chips

For the past four years, the global race to build artificial intelligence has been defined by a single, fragile chokepoint: advanced silicon. The prevailing assumption among Western policymakers and technologists has been that training a frontier-scale AI model requires access to the most sophisticated graphics processing units (GPUs) on the market, almost exclusively designed by Nvidia. If a nation or a company could not acquire that hardware, the logic went, they could not compete at the bleeding edge of AI development. That foundational assumption is now being tested by an unlikely challenger, shifting the geopolitical calculus of the semiconductor industry.[2][6]

The disruption did not come from a state-backed research laboratory or a dedicated AI startup, but from Meituan, a Chinese technology giant best known for its ubiquitous food delivery and local logistics network. On June 30, 2026, the company released LongCat-2.0, a massive open-weight AI model boasting 1.6 trillion parameters. While the sheer size of the model is notable, the parameter count is not the headline. The critical detail is the hardware underneath it: Meituan claims the model was trained entirely from scratch on a cluster of 50,000 domestic Chinese AI accelerators, without utilizing a single Nvidia chip.[1][4]

To understand why this matters, it is necessary to separate the two distinct phases of an AI model's lifecycle: pre-training and inference. Inference is the act of running a finished model to generate answers. It is computationally demanding but relatively straightforward. Earlier this year, other Chinese developers demonstrated the ability to run inference for flagship models using domestic silicon. However, they still relied on stockpiled foreign GPUs for the pre-training phase—the grueling, months-long process of feeding trillions of words into a blank neural network so it can learn the underlying patterns of human language and logic.[1][6]

Pre-training is brutally unforgiving. It requires tens of thousands of chips to operate in perfect synchronization for weeks on end. A single communication timeout, a sudden loss spike, or a silent data corruption event can waste millions of dollars in electricity and compute time. Doing this on a non-Nvidia hardware stack means the challenge is not just about raw processing power; it is about the complex software plumbing required to keep the entire cluster stable. Meituan's achievement suggests that Chinese developers have matured this software infrastructure to a point where it can handle a trillion-parameter training run.[4][6]

The architectural design of LongCat-2.0 is central to how Meituan managed this feat on restricted hardware. The system utilizes a Mixture-of-Experts (MoE) architecture. Instead of activating all 1.6 trillion parameters for every single word it processes, the model routes each task to specialized sub-networks, or "experts." As a result, LongCat-2.0 only activates roughly 48 billion parameters per token. This sparse activation keeps the computational overhead manageable, allowing the model to run efficiently while maintaining a massive overall capacity and a one-million-token context window.[1][4]

The hardware powering this cluster represents a distinct philosophical divergence from Western AI development. While Meituan has not officially named the vendor, industry consensus points to Huawei's Ascend 910C architecture. Due to strict United States export controls enacted in October 2022, Chinese chipmakers have been cut off from the extreme ultraviolet (EUV) lithography machines required to manufacture the world's most advanced transistors. Consequently, domestic chips like the Ascend series are built on older, less efficient manufacturing nodes.[2][3]

To compensate for the limitations of individual chips, Huawei and its partners have leaned heavily into a "scale-out" design philosophy. If they cannot build a single chip that matches an Nvidia H100 in raw performance and power efficiency, they instead focus on networking tens of thousands of slightly weaker chips together. This approach relies on proprietary high-speed fabrics and specialized software, such as Huawei's Compute Architecture for Neural Networks (CANN), to orchestrate massive parallel processing. It is an industrial approach to AI compute, prioritizing predictable scaling behavior over single-chip supremacy.[3][6]

The evidence supporting LongCat-2.0's capabilities extends beyond Meituan's corporate press releases. For two months prior to its official launch, the model was quietly deployed on OpenRouter, a popular developer platform, under the anonymous codename "Owl Alpha." During this stealth period, developers worldwide unknowingly stress-tested the system on complex coding and agentic tasks. The model climbed into the top three by daily usage volume, providing independent validation of its performance before anyone knew it was a Chinese model trained on domestic hardware.[1][4]

Despite this validation, significant uncertainties remain regarding the economics and sustainability of this approach. While Meituan has proven that training a frontier-scale model on domestic silicon is technically possible, the company has not disclosed the financial cost or the power consumption of running a 50,000-chip cluster. The die area and power efficiency of chips manufactured on older nodes are inherently worse than their cutting-edge counterparts. It is highly likely that training LongCat-2.0 required substantially more electricity and physical data center space than a comparable run on Nvidia hardware.[3][6]

Furthermore, the reliability of the cluster during the training run remains a closely guarded secret. While Meituan claims the process was completed without irrecoverable loss spikes, the daily failure rate of individual nodes in a 50,000-chip cluster built on alternative hardware is unknown. If the cluster requires constant maintenance and hardware replacement to stay online, the true cost of this independence may be prohibitive for all but the most well-capitalized technology giants.[4][6]

For policymakers in Washington, the release of LongCat-2.0 forces a reevaluation of the current export control regime. The original intent of the October 2022 restrictions was to freeze China's AI capabilities by denying access to the physical hardware deemed necessary for frontier development. The Meituan case study demonstrates that compute-based containment is not an impenetrable wall. Instead, it has acted as a forcing function, compelling Chinese firms to aggressively optimize their software and networking architectures to route around the hardware deficit.[2][5]

This dynamic suggests that the binding constraint on Chinese AI development has shifted. The question is no longer whether domestic hardware can successfully train a trillion-parameter model, but rather at what financial and environmental cost that training occurs. Export controls are still imposing a severe "friction tax" on Chinese developers, forcing them to solve complex infrastructure problems that their Western counterparts can largely ignore. However, as these software workarounds mature, the effectiveness of hardware-centric embargoes may steadily degrade.[5][6]

For enterprise AI adopters and developers, the geopolitical origins of the hardware are secondary to the model's utility. LongCat-2.0's open-weight release under a permissive MIT license injects another highly capable tool into the open-source ecosystem. Its strong performance in agentic coding tasks, combined with a massive context window, makes it a viable alternative for businesses looking to build autonomous software agents without relying on expensive proprietary APIs.[1][4]

Ultimately, the LongCat-2.0 milestone illustrates the dual nature of technological embargoes. While they successfully delay progress and inflate costs for the targeted nation, they also inadvertently subsidize the creation of parallel, independent supply chains. By proving that the Nvidia ecosystem is not the only viable path to frontier AI, Meituan has not just released a new language model; it has provided a blueprint for hardware independence that other organizations will inevitably attempt to follow.[2][6]