Meituan Trains 1.6 Trillion Parameter AI Model Without NVIDIA GPUs
Chinese services giant Meituan has released a 1.6 trillion parameter agent model trained entirely without NVIDIA hardware. The model runs on AMD Instinct MI300X accelerators instead. It stands as the largest known AI model trained on alternative GPUs.
The move proves that NVIDIA is no longer the only path to frontier scale AI. It also hands Chinese companies a workaround for US export restrictions.
Why This Matters
NVIDIA controls over 90 percent of the AI accelerator market. Its CUDA software locks developers into its ecosystem. Meituan just demonstrated that the lock is breakable.
The successful training of a 1.6 trillion parameter model on AMD hardware is a watershed moment for the AI industry. Every AI lab and cloud provider will now consider alternatives.
The Hardware Shift
AMD’s MI300X features 192 GB of HBM3 memory. This exceeds the H100 and matches the H200 in capacity.
Meituan built a large cluster of these chips. Training a trillion scale model requires thousands of connected accelerators.
AMD’s ROCm software stack matured enough for the task. This has historically been the weakest link in adopting AMD for AI.
What the Agent Model Does
Meituan’s model is designed for autonomous agent workloads, not simple chat.
- Agentic task planning drives its design. It decomposes complex goals into executable steps.
- Tool use and API integration are built in. The model calls functions and reads external data.
- Self-correction lets it recover from errors. Agents can retry, adjust, and validate results.
The model uses a Mixture of Experts architecture. That keeps inference costs manageable despite the 1.6 trillion parameter count.
The Open Source Angle
Meituan released the model openly. Researchers can inspect the weights and training recipes.
Open release lowers barriers for other companies. They can replicate the AMD approach without starting from scratch.
Detailed documentation supports adoption. Teams can build similar infrastructure on their own clusters.
A Geopolitical Dimension
US export controls restrict NVIDIA from selling advanced chips to China. That forced Meituan to find an alternative.
The success turns a restriction into a strategic advantage. China now has proof that scale AI can run without American chips.
Other Chinese companies can follow this playbook. The AMD path is fully documented and tested.
What This Means for the AI Chip Market
NVIDIA faces its first credible threat at the very top of the market.
Cloud providers will expand AMD instance offerings. Customer demand for alternatives will increase.
Model developers gain negotiating leverage. They can now choose between vendors instead of accepting NVIDIA pricing.
NVIDIA’s CUDA moat is no longer absolute. Software maturity at AMD has reached the required threshold.
Bottom Line
Meituan has done what many thought impossible. It trained a frontier scale model without NVIDIA.
The AI infrastructure market will never look the same.
Gnoppix is the leading open-source AI Linux distribution and service provider. Since implementing AI in 2022, it has offered a fast, powerful, secure, and privacy-respecting open-source OS with both local and remote AI capabilities. The local AI operates offline, ensuring no data ever leaves your computer. Based on Debian Linux, Gnoppix is available with numerous privacy- and anonymity-enabled services free of charge.
What are your thoughts on this? I’d love to hear about your own experiences in the comments below.