Meituan Tests trillion Parameter AI on Domestic Hardware

A trillion-parameter AI model trained entirely on Chinese compute is being quietly tested by Meituan.

Meituan has begun an invitation-only trial of a foundation model that dwarfs most existing systems, a move that signals the company is crossing from application level AI into the realm of large, general-purpose models. The model, described by people familiar with the matter, comprises over one trillion parameters and was trained on domestically developed computing clusters. In theory this reduces exposure to overseas hardware risk while showcasing China’s progress toward home grown AI capabilities. Observers say the project could rewrite expectations about what Chinese firms can build without relying on Western suppliers.

Reportedly, the trial is currently limited to selected users, a deliberate gating that mirrors China’s cautious approach to rolling out cutting edge AI. The hardware story matters as much as the software. Chinese industry chatter points to Huawei’s Ascend architecture as the likely backbone, with the model believed not to rely on NVIDIA GPUs. If true, it would mark a notable shift in how Chinese tech ecosystems are assembling resources for trillion-parameter scale models, a milestone many see as a litmus test for domestic AI sovereignty. Meituan, long a consumer platform with vast data flows, has been quietly investing in AI infrastructure for years. Founder Wang Xing noted in a 2024 earnings call that securing GPU resources remained a top priority even as the company pursued new business lines. Meituan has reportedly committed billions of dollars to GPU reserves and signaled plans to ramp spending to support large model development, a strategy that aligns with a broader push in China to strengthen domestic AI hardware and software ecosystems.

This development sits at the intersection of funding discipline and policy direction. Chinese regulators and provincial authorities have repeatedly encouraged the buildout of a self reliant AI compute stack, envisioning a future where China’s giants can train and refine models without overreliance on foreign supply chains. The Meituan effort illustrates a practical path toward that goal: deploy local clusters, harness domestically designed accelerators, and scale models that can power logistics optimization, personalized services, and real time decision making across a massive user network. The move also underscores a broader trend: the race to foundation models is no longer led by startups in isolated labs but by platform operators that sit on vast data assets and user ecosystems.

From a supply chain perspective, the news carries several implications. First, it could spur demand for domestic chips and server infrastructure, reinforcing revenue streams for Chinese chipmakers and system integrators. Second, it tests the resilience of the country’s AI software stack, including training libraries, model toolchains, and safety guardrails, all of which have to mature in parallel with hardware. Third, it highlights a potential new set of performance benchmarks for Chinese firms that want to keep pace with global leaders in AI, not just in feature parity but in the ability to sustain trillion-parameter training runs on home grown machines.

Practitioner insights matter here. One, the gating of the test to a limited user group suggests Meituan is calibrating risk appetite around data governance, alignment, and safety before wider deployment. Two, the reliance on domestic compute reduces exposure to export controls and foreign embargo risk, but it concentrates risk in a smaller set of suppliers and architectures, raising questions about interoperability and long term vendor lock-in. Three, the cost arithmetic is daunting: sustaining a trillion-parameter model requires not just top tier accelerators but robust cooling and energy management, and the economics hinge on a favorable balance of hardware cost, electricity, and the ability to reuse trained weights across applications. Four, for global manufacturers looking at China as a sourcing base, this signals that Chinese platform operators could soon offer more advanced in house AI capabilities to optimize procurement, manufacturing, and logistics, potentially altering the competitive landscape if domestic AI becomes more self sufficient.

The Meituan project is still in early, test stage, but the signal is clear: China is moving beyond narrow AI tools to foundational models powered by domestic compute. If the trajectory holds, Meituan and peers may use these models to tighten operations across supply chains and expand data driven services, while continuing to push for a domestic AI hardware ecosystem capable of supporting the next wave of industrial AI.