Fireworks AI cofounder: training is now an economics play for AI apps
sophiamyang · x · 2026-11-05
Fireworks AI cofounder thebunnychen on why app companies train custom models:
- Training is now an economics play: sustained usage costs, not acquisition costs, limit AI app scale — the inverse of Web 2.0 where serving an extra user was nearly free and ads drove costs.
- Cheaper models are a growth enabler: cheaper → more users → more data and better evals → faster scaling. Founders optimize wall-clock time and will pay more for faster, better results.
- Fireworks reuses training recipes across customers so startups don't repeat each other's mistakes.
- "No eval" approach: teams can start training without offline evals — your traffic is the training signal.
More from Companies & People
- Cosmos and Inherent Labs launch £20,000 essay prize on AI and human autonomy — sebkrier · 2026-11-05
- Matthew Berman quips lab-hopping AI researchers are making 'generational money' — MatthewBerman · 2026-11-05
- Runway co-CEO's origin story: taught himself coding at 10, now $200M ARR — Scobleizer · 2026-11-05
- AI engineering educator jxnl.co announces he has joined OpenAI — jxnlco · 2026-11-05
- So close: user misses X creator payout threshold at 496.6K of 500K impressions — alexmacgregor__ · 2026-11-05
- Alexandr Wang says 'reporting for duty' as Charlie Holtz and others respond — charlieholtz · 2026-11-05