One company spent $500 million on AI in a single month. Uber burned through its entire annual AI budget in just four months. Microsoft quietly pulled 100,000 Claude Code licenses.
The enterprise AI cost crisis isn't coming — it's already here. And today, June 1st, GitHub Copilot officially switches to usage-based billing, making the new reality official for millions of developers.
The Problem: AI Spending Is a Black Hole
Here's the uncomfortable truth: 73% of AI projects show zero measurable ROI, according to a new MIT study analyzing $665 billion in enterprise AI spending. Nearly half — 48% — of executives now call their AI adoption a "massive disappointment."
The math doesn't work. Companies bought into the vision of AI transforming everything, handed out licenses by the thousands, and watched token costs multiply faster than anyone budgeted for.
Microsoft's move to pull 100K licenses wasn't a strategic pivot. It was a financial correction. When you're paying per-token and developers discover that AI agents can spin through your budget in hours, the economics collapse fast.
Per-task token consumption is growing faster than per-token cost drops. Translation: even as models get cheaper per token, agents are using so many more tokens per task that the total bill keeps climbing.
The Solution: Metered AI and Token Governance
The industry is responding with metered pricing — pay for what you use, with hard caps and accountability built in.
GitHub Copilot's shift to usage-based billing (effective today) is the canary in the coal mine. Anthropic follows on June 15th with separate agent credit pools. This isn't a bug — it's the new business model.
What works:
- Per-team token budgets with automatic throttling when limits hit
- Outcome-based pricing models that tie AI spend to actual business results
- Usage governance dashboards that show which teams, projects, and agents are burning cash
- Agent-level cost attribution — knowing exactly which workflow ate $50K last Tuesday
The companies surviving this transition share one trait: they treat AI spend like cloud spend. Measured, governed, and tied to outcomes.
The Numbers
- $665 billion — total enterprise AI spending analyzed by MIT
- 73% — ROI failure rate across those investments
- 48% — executives calling AI adoption a "massive disappointment"
- 95% — AI pilots delivering zero measurable P&L impact
- 100,000 — Claude Code licenses Microsoft pulled back
- $500M — single company's monthly AI burn (reported, not named)
- 4 months — time it took Uber to exhaust its annual AI budget
- June 1, 2026 — GitHub Copilot switches to usage-based billing
- June 15, 2026 — Anthropic launches separate agent credit pools
Caveat: The $500M figure comes from industry reports and hasn't been independently audited. The MIT study covers a broad definition of "AI spending" that includes infrastructure, not just API calls. But the directional signal is clear — costs are out of control at scale.
The Impact
This isn't just a finance problem. It's a survival problem for companies building on AI.
For CTOs: If you can't tell your CFO exactly what each AI agent costs per completed task, you're flying blind. The companies that build cost observability now will be the ones that survive the metered era.
For AI-first startups: Your unit economics just changed. If your product relies on heavy LLM usage, your cost structure needs to be as rigorous as a SaaS company tracking AWS spend. Token costs are the new cloud bill.
For the industry: The shift from unlimited to metered isn't a step back — it's maturity. We went through the same cycle with cloud computing. The "free tier" era ended, and the companies that built cost discipline early won.
The $965 billion Anthropic valuation and $47 billion run-rate revenue announced this week? That money is coming from somewhere. It's coming from enterprise budgets that are finally being forced to confront what AI actually costs.
The Bottom Line
The era of "just give everyone AI and see what happens" is dead. It died when the bills came in.
If your company doesn't have token budgets, usage governance, and cost-per-outcome metrics in place by Q3 2026, you're going to be the next case study in how not to adopt AI. The tools exist. The frameworks exist. The data exists.
What's missing, for most companies, is the discipline to use them.
Stop asking "how do we get more AI?" and start asking "what did the AI we already have actually deliver?" That single question will save you more money than any model optimization ever could.