Analysis
Meituan open-sourced LongCat-2.0, a 1.6-trillion-parameter agentic coding model trained entirely on Chinese-made chips, according to VentureBeat. The release is significant less for its benchmark scores -- described as near-frontier -- than for what it proves about domestic chip capability: that Chinese AI infrastructure can now train models at genuinely large scale without relying on the advanced Western AI chips subject to export restrictions.
The context is years of escalating U.S. export controls aimed at limiting China's access to the most advanced AI training chips, a policy predicated partly on the assumption that domestic Chinese chip alternatives would meaningfully lag Nvidia and other Western silicon for large-scale model training. LongCat-2.0's training run is a direct data point against that assumption -- a 1.6 trillion-parameter model, among the larger openly disclosed model sizes, trained without foreign chips.
The strategic logic for Meituan and the broader Chinese AI ecosystem is twofold: demonstrating self-sufficiency reduces geopolitical vulnerability to further export restrictions, and open-sourcing the resulting model maximizes global mindshare and adoption despite -- or because of -- the chip constraints. This mirrors DeepSeek's earlier strategy of using open releases to build developer trust and usage regardless of the underlying compute story.
“This mirrors DeepSeek's earlier strategy of using open releases to build developer trust and usage regardless of the underlying compute story.”
The competitive landscape now includes multiple Chinese labs -- DeepSeek, Meituan, and others -- racing to prove that domestic chip ecosystems can support frontier-adjacent model development, while Western labs including Anthropic, OpenAI and Google continue to rely heavily on Nvidia's most advanced chips (and now increasingly custom silicon from companies like Etched). The bifurcation between chip ecosystems is becoming a defining structural feature of the global AI industry.
For founders and investors, the practical implication is that a widening set of near-frontier, freely available models means capability commoditization is arriving faster and from more directions than pure API pricing wars alone suggest -- open weights from Chinese labs are increasingly a viable foundation for products that don't require the absolute frontier.
The bear case is that 'trained entirely on Chinese chips' claims are difficult to independently verify, and near-frontier benchmark performance doesn't necessarily translate into production-ready reliability or safety guarantees Western enterprise buyers require. What to watch: independent benchmarking of LongCat-2.0 against Western frontier and near-frontier models, adoption rates outside China, and whether the training-hardware claim holds up to scrutiny.