Why it matters
If routing to cheaper models grows usage instead of curbing it, the durable winners are Microsoft, Amazon, and Google, who bill for compute no matter which model answers. Model quality stops being the moat; owning the infrastructure is.
Tokenmaxxing read
Jarc's tollbooth image: every routed token pays the same cloud toll no matter how cheap the model. Microsoft says it processed over 100 trillion tokens in one quarter of 2025, up 5x year over year — so routing work to small models still swells the bill hyperscalers collect.
Source takeaway
Rihard Jarc (UncoverAlpha, June 2026): inference for GPT-3.5-level quality dropped more than 280x in two years, so cheaper tokens unleashed far more of them. Frontier labs run up to ~70% gross margins; cloud infra sits near 35% — but captures the volume.
