In its latest earnings call, Pinterest stated that in its actual business, such as Pinterest Assistant, the company's use of open-source models combined with its own proprietary data for post-training has resulted in performance superior to third-party closed-source models in certain scenarios. More importantly, the cost per transaction is less than 8% of comparable closed-source proprietary models.
This case illustrates that the decline in enterprise AI costs does not necessarily depend on a drop in the price of underlying GPUs or computing power, but may stem from changes in model selection and customization methods. Enterprises can use lower-cost open-source models and then post-train them using their own data, rather than having all tasks call more expensive, cutting-edge closed-source models. This aligns with the recent trend of enterprises reducing token expenditures through intelligent routing: with an increasing number of AI models, what enterprises truly need to optimize is "which model is worth using for which task." If this model expands, the volume of AI applications can continue to grow, while the unit inference cost may not increase proportionally.
