China’s Meituan says new AI model trained on domestic chips

Meituan said it released LongCat-2.0, a claimed trillion-parameter LLM trained and run on a 50,000-chip Chinese cluster using domestically powered processors, and will open-source the model. The company positioned it for “agentic” commerce and coding, citing up to 1 million-token input handling and benchmark parity/exceeding proprietary models on some tasks. While integration details weren’t disclosed, the open-source launch and self-sufficiency narrative around U.S. export controls may be supportive for sentiment around China’s domestic AI ecosystem.

Analysis

This is less about model quality and more about strategic capex leverage: if a consumer internet platform can credibly train at scale on domestic silicon, the market should assign a higher survival multiple to China tech because the dependency on U.S. accelerators is no longer binary. The immediate winner is the domestic AI supply chain; the listed proxy we can actually own is BABA, where the signal is that agentic commerce is moving from a feature to a margin-defense tool for merchant engagement and conversion.

Second-order, open-sourcing matters because it seeds standards. That tends to widen the moat for the ecosystem that can provide chips, cloud, and distribution together, while compressing pricing power for standalone model vendors over 6-18 months. For GOOGL, the direct earnings impact is small, but the broader takeaway is that frontier-model scarcity is eroding faster than the market likely expects, which can shave some multiple support from the AI complex if investors start treating model access as a commodity.

The main risk is verification: benchmark claims and “trained on domestic chips” are low-conviction until we see product usage, latency, and cost savings in live commerce flows. Over the next 1-3 months, watch for disclosure around AI-driven take rates, merchant adoption, or cloud/compute mix; if those don’t show up, this is mostly narrative. Falsifier: if BABA/China internet margins do not improve by the next two earnings cycles, or if domestic-chip training proves too expensive to scale, the self-sufficiency premium should fade quickly.

Key Decisions for Investors

• Buy BABA on a 3-5% pullback or via a 3-6 month call spread; thesis is that AI-assisted commerce becomes a margin-defense lever, with upside if the market starts capitalizing China internet AI optionality more like cloud infrastructure than e-commerce.
• Relative-value: long BABA vs. short GOOGL in a small, hedged ratio if the market overreacts to Chinese AI self-sufficiency headlines; expected holding period 1-3 months, with the trade invalidated if GOOGL re-rates higher on accelerating AI monetization or if BABA fails to show product-level adoption.
• Do not chase GAP/TGT on this headline; keep U.S. consumer discretionary exposure unchanged unless there is evidence of a broader China demand shock or tariff response. This is a watch item, not a spillover trade.