Article
Analysis: Domestic large models enter an "awkward phase": training compute, inference costs, and pricing become major challenges
Beating AI flash news: Max For AI posted that domestic large models are currently entering an "awkward phase," with core challenges concentrated in training compute, inference resources, and commercial pricing.
It stated that as domestic vendors continue to expand model scale, parameter scales of,, and even have begun to be discussed, but growth in training compute may struggle to keep pace with expanding model scale, with compute gradually becoming a practical constraint. Alibaba previously mentioned at the Apsara Conference that it may train models with parameter scales of to in the future, while Kimi and GLM are also continuing to advance large-scale models.
On the inference side, completing large-model training does not mean services can be provided reliably to large numbers of users over the long term. Max For AI stated that some large models run relatively quickly when first launched, but may subsequently experience throttling, queues, and tighter usage limits, reflecting that inference compute remains a significant bottleneck. As model scale expands, each additional increment of user quota entails higher actual inference costs.
In addition, rising model scale and inference costs may force vendors to adjust prices, while open source and cost-effectiveness have been among the key competitive advantages of domestic large models. It believes that in the Agent era, as per-task consumption Token increases, price sensitivity may rise further, and domestic vendors will need to rebalance model performance, inference costs, and commercial pricing.
Original link https://m.theblockbeats.info/flash/369003