The parameter scale of large models has increased by 100 times, now surpassing the trillion-level threshold, resulting in significant resource consumption and escalating costs for storage, inference, operations, and implementation. Large model enterprises are actively engaging in a "cost-trimming" movement. Firstly, data is being scaled up to enhance the marginal benefits of data through economies of scale; secondly, models are being compressed to operate with faster inference speeds, lower latency, and reduced resource requirements without compromising performance; thirdly, computational efficiency is being improved by enhancing the performance of chips and computing clusters; fourthly, business stratification is occurring, with distinct commercial paths emerging for large models of varying sizes, functionalities, and orientations. To ensure long-term, sustainable service, large models must undergo "cost-trimming," a necessary journey.