20VC: Fireworks, Open-Source AI & Why Token Costs Could Drop 10X

The future of artificial intelligence lies in decentralized, specialized intelligence rather than a single universal model. Lin Qiao, founder of Fireworks and former systems architect at LinkedIn and Facebook, argues that because the majority of valuable data is locked within private enterprise silos, intelligence must be distributed across many models tailored to specific company workflows. This shift from "one size fits all" to "one size fits one" necessitates a focus on the inference layer, where businesses customize open models to maintain control over weights and data. Economic sustainability in this new era depends on a projected 10x reduction in token costs over three years, which is expected to trigger a 100x explosion in usage. As Fireworks processes over 40 trillion tokens daily, the industry moves toward multi-model systems where specialized models handle distinct tasks like legal work, coding, or recruiting, decoupling product-market fit from the challenges of large-scale deployment.