China's artificial intelligence ecosystem has executed a landmark maneuver that could reshape the global semiconductor industry landscape. Meituan, a company primarily known as a giant in the delivery services sector, has unveiled its flagship language model, LongCat-2.0. However, the main sensation lies not in its size (1.6 trillion parameters), but in how it was created.
LongCat-2.0 has become the first model of this scale in China to undergo a complete training cycle — from pre-training to inference — exclusively on domestic equipment. While most Chinese developers, including DeepSeek, used local chips only for the inference stage, Meituan went further. They utilized a cluster of 50,000 proprietary ASIC accelerators built on an architecture different from Nvidia's graphics processing units.
Architectural Breakthrough and the "Bottleneck"
The key difference was the use of specialized ASIC chips designed for specific tasks. The company did not disclose the direct supplier of the hardware, but confirmed the use of Huawei's communication library (HCCL) to stabilize the process. This indicates that the infrastructure is likely based on Huawei Atlas 950 clusters, which, according to analyst TP Huang, now leave no doubt about their ability to train large language models.
However, there were compromises. The main bottleneck, as acknowledged by Meituan itself, was memory. Chinese ASIC accelerators have significantly less memory per device compared to the Nvidia H800 chip, which is banned for export to China. To compensate for this shortfall, engineers had to make tremendous efforts to create a stable and scalable infrastructure. This demonstrates that circumventing sanctions requires not just replacing the hardware, but a deep overhaul of the software and system stack.
Benchmarks and the Real Situation
LongCat-2.0, with a context window of 1 million tokens, has already shown impressive results in programming and agent tasks, surpassing Google's Gemini 3.1 Pro in Terminal-Bench 2.1 and SWE-Bench Pro tests. Nevertheless, Meituan honestly admits that the model still lags behind the flagships of OpenAI (GPT-5.5) and Anthropic (Claude 4.8 Opus) in a number of general tests, and has not been evaluated on leading independent platforms.
My professional opinion: Meituan's success is not just a technical victory, but a powerful signal for the market. It proves that the era of Nvidia's "total dominance" in model training is coming to an end. China has created a working precedent that, within the next 12-18 months, will trigger explosive growth in demand for domestic ASIC solutions, which in turn will put pressure on GPU prices and accelerate the diversification of supply chains worldwide. For the crypto industry, where access to computing power is critical, this could mean the emergence of new, cheaper alternatives for mining and data processing.