KEY POINTS
- Fireworks AI VP Rob Ferguson says most companies are better served by managed inference than by running their own GPU datacenters.
- He claims routing work across a portfolio of open and closed models beats relying on any single model.
- Ferguson says customers should track cost per task rather than token price, and claims Fireworks can cut a mixed-model bill by half.
The head of a managed inference company is urging businesses to stop leaning on a single AI model and to avoid building their own GPU infrastructure where possible. Rob Ferguson, a vice president at Fireworks AI, made the case in an interview with The Register, which published the conversation on 7 November 2026.
