Why Smarter AI Routing Is the Cost Advantage Nobody Is Talking About
1 min read
Originally from youtube.com
View source
My notes
Summary
Eric Glyman (CEO of Ramp) frames the AI cost curve as bifurcated: basic AI tasks are 700x cheaper while frontier tasks are 300x more expensive. The arbitrage is smart token routing, sending low-complexity work to cheap models and only escalating to frontier models when needed. Enterprises are already asking vendors for help with “token optimization”.
Key Insight
- The cost gap between basic and frontier model tiers is widening, not closing. Using GPT-4-class for trivial tasks is a 700x overpay.
- “Token routing” is becoming a discrete product/skill category, large enterprises are reaching out to peers in CEO chats asking for help.
- The framing “I don’t need the best intelligence to clean a toilet” is a useful mental model for client conversations: most agentic workflows have 80% janitor steps and 20% brain steps.
- Implication: building anything LLM-powered today without a router layer leaves margin on the table.