Domestic Chips Train Trillion-Parameter Model, Highlighting China's AI De-Americanization

1. The “Chinese Solution” for Trillion-Parameter Models

DeepSeek V4, released only months after its predecessor, is a trillion-parameter model that is fully open-source, challenging the dominance of closed-source offerings and showing that Chinese algorithms can compete internationally.

At the same time, Meituan has successfully trained a model of comparable scale using a compute cluster built entirely from domestically produced chips, proving that the hardware side of large-model training is no longer limited to NVIDIA GPUs.

2. Meituan’s “Hidden Battlefield”

Beyond its well-known food-delivery business, Meituan relies heavily on AI for tasks such as intelligent dispatch, autonomous delivery, recommendation, and voice interaction, creating a massive and rigid demand for advanced models.

By choosing an all-domestic compute cluster, Meituan signals that domestic chips now deliver sufficient performance for the most demanding training workloads, and that the company is strategically positioning itself ahead of supply-chain uncertainties.

The open-source nature of DeepSeek V4 allows Meituan to customize and privately deploy the model, offering data-security advantages that closed-source alternatives cannot match.

3. Resonance Between Open-Source and Domesticization

Open-sourcing DeepSeek V4 is not merely charitable; it accelerates ecosystem growth by inviting developers to contribute, speeding bug fixes, performance improvements, and breaking the “black-box” monopoly of large-model providers.

Meituan provides a concrete deployment scenario—trillion-parameter model + domestic compute + real-world business needs—demonstrating a reproducible paradigm that other enterprises can follow to run large models on Chinese chips.

This approach is presented as a genuine second option rather than a backup.

4. Trend Judgment: China’s Dual-Track AI Roadmap

Combining DeepSeek V4’s open-source breakthrough with Meituan’s domestic-chip training illustrates a clear trajectory: Chinese large-model development is advancing along two parallel tracks, one focused on algorithmic innovation and the other on self-controlled compute infrastructure.

DeepSeek V4 proves “we can build,” while Meituan proves “we can use.” The next phase involves more companies joining this ecosystem, moving Chinese models from merely usable to truly effective, with trillion-parameter scale serving as a starting point rather than an endpoint.

When open-source ecosystems and domestic chips form a positive feedback loop, China’s AI foundation becomes independent of external forces, which is the most significant implication of the coincident developments.