Fireworks CEO: Specialized AI, Not General, Wins Enterprise

Fireworks CEO Lin Qiao asserts the company processes over 40 trillion tokens daily, surpassing OpenAI and Gemini APIs. She positions this as validation for a world where intelligence is not dominated by frontier labs but by millions of specialized models per application. The core thesis is that private enterprise data, not public internet data, will

drive the next AI frontier, requiring continuous fine-tuning and ownership of model weights. Qiao argues that fine-tuning and RL are essential moats for companies, as the barrier to building AI features is now thin. She warns of 'scaling to bankruptcy' due to expensive AI inference, stating that cost control and model customization can yield 5-6x

savings per task compared to closed APIs. This is presented as a necessity for durable business models, shifting the focus from token maximization to value maximization. Counterintuitively, Fireworks relies heavily on open-source models but keeps its inference and training engine proprietary, citing the need for extreme optimization and rapid iteration. The team of about 10

engineers manages over 100,000 configuration options per workload. Qiao refuses to open-source the engine, arguing it would confuse a fast-moving community and that supporting existing open-source projects like vLLM is more productive. For engineering leaders, the key takeaway is that the current AI inflection point is not just about model capability but about operational and