According to CoinMeta, analysts max, for, and ai stated that domestic large models are currently entering an "embarrassing period," with core challenges focusing on training computing power, inference resources, and commercial pricing. As domestic manufacturers continue to expand the scale of their models, model sizes of 5T, 8T, and even 10T are being discussed, but the growth rate of training computing power may struggle to keep up with this expansion, making computing power a gradual limiting factor. Alibaba previously mentioned at the Yunqi Conference that it might train models with 5T to 10T parameter sizes in the future, and manufacturers such as kimi and glm are also continuing to advance large-scale models. On the inference side, the completion of large model training does not necessarily mean that they can provide stable services to a large number of users in the long term. Some large models operate quickly in the initial stages after going live, but later on, issues such as speed limitations, queuing, and tightened quotas may arise, indicating that inference computing power remains a significant bottleneck. As the model size increases, each additional user quota comes with higher actual inference costs. Furthermore, the rise in model size and inference costs may force manufacturers to adjust their prices, whereas one of the key competitive advantages of domestic large models before was open-source development and cost-effectiveness.