One company is burning through $500 million a month on Claude API calls. Uber blew their entire 2026 AI budget by April. Microsoft had to revoke Claude Code licenses across the company because costs spiraled so fast.
This isn't a pricing problem. This is a fundamental mismatch between how companies budget for AI and how AI agents actually consume resources.
The Problem: Per-Token Pricing Met Enterprise Agents
Here's the math that's breaking finance departments: per-token pricing dropped 98% since 2023. Sounds great, right? Except agentic AI consumption grew 300x in the same window.
A single coding agent running autonomously for 12 hours — which Claude now does routinely, up from 30 minutes just 18 months ago — can burn through more tokens than an entire department used in a month of ChatGPT queries.
The Sinch survey published this week put hard numbers behind what everyone's feeling: 74% of AI agents get rolled back from production, and one of the top three reasons is budget loops. Not model quality. Not accuracy. The agent literally spent too much money doing its job.
The Bain survey adds another layer: 40% of companies are seeing less than 10% cost savings from AI, against targets of 11-20%. And here's the vicious cycle — 44% are funding their next wave of AI investment from "savings" that never materialized. It's AI budgets all the way down.
The Solution: Token Economics Is Now a Core Engineering Discipline
The Tokenomics Foundation launches in July, which tells you everything about how urgent this has become. But you don't need to wait for a foundation to get costs under control.
Here's what actually works:
Budget fencing. Every agent gets a hard spending cap per task, per day, per project. Not a soft alert — a hard kill switch. If the agent hits the ceiling, it stops. Period. No "just one more retry."
Token budgeting as a first-class metric. Teams shipping AI features need to track token cost per task the same way they track latency and error rates. Microsoft's move to usage-based Copilot billing isn't random — it's the new normal.
Model routing. Not every task needs Claude Opus. Route simple classification to a smaller model. Reserve frontier models for complex reasoning. NVIDIA's Nemotron 3 Ultra (550B MoE, only 55B active per inference) is purpose-built for this — aggressive token efficiency for long-running agents.
Human-in-the-loop cost gates. The most consistently exploited failure mode in Microsoft's new Agentic Failure Taxonomy? Bypassing human approval. If your agent can spend money without a human signing off, you've already lost.
The Benchmarks: What the Numbers Actually Look Like
- $500M/month — confirmed single-company Claude bill (enterprise, heavy agentic use)
- Uber — entire 2026 AI budget exhausted by April 2026
- 98% — per-token price reduction since 2023
- 300x — agentic token consumption growth in the same period
- 74% — AI agent production rollback rate (budget loops among top 3 causes)
- 40% — companies seeing fewer than 10% AI savings vs 11-20% targets
- 44% — funding next AI wave from prior "savings" that underdelivered
- 90% — companies increasing AI budgets anyway despite underwhelming returns
Caveat: The $500M figure comes from industry reporting and hasn't been independently verified by public financial filings. Uber's budget exhaustion is reported, not from Uber's official statements. Take individual numbers with appropriate skepticism — but the directional trend is undeniable.
The Impact: What This Means for Your Business
If you're deploying AI agents without token-level cost tracking, you're flying blind. Full stop.
The companies surviving this crisis treat AI spend like cloud spend circa 2015 — something to be obsessively monitored, optimized, and architectured for. The ones struggling treat it like a SaaS subscription they can set and forget.
For a mid-market company running 50 agents across customer support, sales, and engineering, the difference between uncontrolled and optimized token spending is easily $200K-$500K per month. That's the difference between "AI is transforming our business" and "we need to shut down the AI program."
The ROI gap Bain identified isn't a technology problem. It's a measurement problem. You can't optimize what you don't track, and most companies have zero visibility into which agents are productive and which are burning tokens in loops.
Here's my take: per-token pricing will be remembered as a transitional model. Within 18 months, most enterprise AI will move to outcome-based or capacity-based pricing. The Tokenomics Foundation is the canary in the coal mine. If you're not building cost observability into your AI stack right now, you're building technical debt that will be excruciating to retrofit.
The companies that figure out token economics first won't just save money — they'll be the only ones who can afford to run the agents everyone else is rolling back.