Lin Qiao explains that Fireworks AI focused on inference because most valuable enterprise data is private and not used for training general models. She champions open-source models, which now match closed models in quality and offer cost-effective customization. She predicts token costs will drop 10x in three years, driving 100x usage growth, and envisions a multi-model world where tasks are routed to specialized models. Qiao identifies the lack of a system designed for very large models as the key bottleneck, and argues that every company will eventually own its own intelligence, similar to custom software stacks. She also notes that Fireworks AI processes over 40 trillion tokens daily, mostly from customized models, and sees potential for 20-100x growth.