Descriptions:
Lin Qiao, founder and CEO of Fireworks AI, sits down with Harry Stebbings on 20VC for a wide-ranging conversation about the future of AI inference, open-source intelligence, and the business of building specialized AI infrastructure. Qiao, who started Fireworks at age 48 after senior roles at Facebook and LinkedIn, argues that general-purpose models trained on public internet data represent only a fraction of the world’s intelligence opportunity — the real prize lies in activating private enterprise data that will never reach a public training corpus.
The conversation covers Fireworks’ technical approach to distributed reinforcement learning, including their partnership with Cursor on decoupling the RL trainer and rollout processes across multiple global data center regions. Qiao explains how this architecture reduces dependence on expensive interconnected GPU clusters while keeping model weights fresh enough to maintain training signal quality — an innovation he says enabled Cursor’s recent model launches.
On market outlook, Qiao predicts a 10x reduction in token costs over the next three years, which he believes will catalyze a 100x expansion in AI usage. He also draws a sharp distinction between what he calls “last year’s” coding wave and the current shift toward AI co-work, describing co-work as a more diversified and expansive opportunity. He is unequivocal that Fireworks will not move into the application layer, though infrastructure and data center investments remain under consideration.
📺 Source: 20VC with Harry Stebbings · Published July 20, 2026
🏷️ Format: Interview
