Most AI Startups Are Scaling Into Bankruptcy | Lin Qiao, CEO of Fireworks

Specialized intelligence offers a more efficient and proprietary alternative to general-purpose models for enterprise applications. By focusing on customized models rather than black-box APIs, companies retain their unique data and intellectual property, creating a sustainable competitive moat. Fireworks, led by founder and former PyTorch expert Lin Qiao, processes over 40 trillion tokens daily, demonstrating that specialized models deliver superior performance at five to six times lower costs than general-purpose counterparts. The industry is currently shifting from "token maxing" toward "value maxing," prioritizing ROI and domain-specific reasoning over raw scale. This approach enables businesses to integrate proprietary knowledge directly into their models, ensuring that intelligence remains an internal asset rather than an outsourced commodity. Continuous, iterative fine-tuning and reinforcement learning allow these specialized models to evolve alongside rapidly changing application needs.