Analysis: Domestic large models enter an 'awkward period' – training computing power, inference costs, and prices have become three major challenges.

Beating AI Express News: Max For AI stated that domestic large models are currently entering an "awkward period", with core challenges focusing on three areas: training computing power, inference resources, and commercial pricing. As domestic vendors continue to expand model scales, sizes of 5 trillion, 8 trillion, and even 10 trillion parameters are being discussed, but the growth rate of training computing power may struggle to keep pace with this expansion, making computing power a practical constraint. Alibaba previously noted at its Cloud Summit that it may train models with 5T to 10T parameters, while vendors including Kimi and GLM are also advancing large-scale model development. On the inference side, completing large model training does not equate to being able to stably serve a large user base long-term. Max For AI pointed out that some large models run fast at launch, but may later face speed limits, queuing, and tightened quotas, reflecting that inference computing remains a key bottleneck. As model scales grow, each additional user quota translates to higher actual inference costs. Additionally, rising model sizes and inference costs may force vendors to adjust pricing, as one of the key competitive advantages of domestic large models has been open-source availability and cost-effectiveness. The firm added that in the Agent era, as single tasks consume more tokens, price sensitivity will further increase, requiring domestic vendors to rebalance model performance, inference costs, and commercial pricing.