Cloud taught two generations of engineers to fear the monthly invoice. AI is teaching a third.
OpenRouter's public stats show hundreds of billions of tokens routed through their platform alone. Add Groq, Together, Fireworks, Azure, Bedrock and the rest, and the industry number lands north of 50 trillion tokens a day. That's tokens someone paid for.
What's different: cloud bills were predictable by instance. AI bills vary by prompt complexity, retry logic, and agent loops. A single runaway agent can eat a quarter's budget in an afternoon.
What good teams do now: token budgets per feature, alerting on spend vs. forecast, model routing by task complexity (big model only when needed), and a hard ceiling on agent loops.
The tooling catching up: OpenRouter's spend caps, Helicone's cost analytics, Portkey's automatic routing. If your AI feature doesn't have a cost guardrail, you don't have an AI feature — you have a roulette wheel.
This is why open-weight + local inference keeps winning more share: predictable economics, zero per-token risk, and the ability to cap costs at the GPU line item.