Meituan Trains 1.6T Parameter Model Without Nvidia, Signaling China’s AI Hardware Shift

Published by James Harris on June 30, 2026June 30, 2026

Meituan, better known for delivering dumplings than deploying frontier AI, claims its new LongCat-2.0 model reaches 1.6 trillion parameters without touching a single Nvidia GPU. The company says it trained the model entirely on domestic Chinese ASIC superpods, which, if accurate, makes it one of the more concrete examples of a large Chinese firm operationalizing a post-Nvidia stack at scale.

The timing is not coincidental. U.S. export controls have progressively narrowed the hardware options available to Chinese AI developers, and each new restriction has accelerated domestic alternatives rather than simply slowing development. Bernstein estimated Nvidia still holds roughly 40 percent of China’s AI chip market in 2025, with Huawei at a similar share, but projected Nvidia losing around eight percentage points this year as Huawei gains ground. LongCat-2.0’s sparse mixture-of-experts architecture, the same broad approach used by DeepSeek and Mistral’s Mixtral, activates only a subset of parameters per token rather than the full 1.6 trillion, keeping inference costs manageable at a scale that would otherwise be prohibitive. Its one million token context window is also a meaningful spec gap over DeepSeek-R1-0528 and GPT-OSS, both capped at 128,000 tokens, though Meituan’s own benchmarks against Google, OpenAI, and Anthropic have not been independently verified.

Self-reported benchmarks from a company with no prior AI credibility are exactly as reliable as they sound.

The more consequential question is portability. Meituan says the core reasoning architecture can run beyond domestic hardware, but optimization for Chinese ASICs may create friction on Nvidia-based infrastructure, which still dominates data centers outside China. That limits near-term developer adoption internationally, regardless of what the benchmarks show. Meituan entered AI seriously only after acquiring Light Year Beyond for $281 million in 2023 and did not publicly announce internal model development until 2025, so the institutional AI muscle here is still shallow. MiniMax, a better-established Chinese AI startup backed by Alibaba, rolled out its own million-token-context model in early June 2026 at pricing well below U.S. market leaders, which means Meituan is entering a domestic competitive environment that is already compressing margins.

What LongCat-2.0 actually demonstrates, independent of its benchmark claims, is that U.S. export controls are producing a parallel hardware ecosystem faster than most analysts expected two years ago. Whether that ecosystem produces models competitive at a global level is a separate question, and one that only third-party evaluation can answer.