The all-you-can-eat AI buffet just closed, and finance teams are staring at bills that would make a data center blush.
The Summary
- Coinbase instituted weekly AI spending caps ranging from $500 to $5,000 per employee after usage went "parabolic" following Anthropic's Claude Opus 4.6 launch in February
- Companies like Deloitte are also testing new limits on AI spending as prices for popular AI tools have risen, creating enterprise sticker shock
- The shift marks a transition from "are people using AI?" to "are people using it well?"—the first real optimization phase of enterprise AI adoption
The Signal
Winter 2026 was the season of AI gluttony. Companies told employees to use AI coding assistants without limits, to experiment, to reimagine workflows. Then the bills arrived. Coinbase executive Rob Witoff watched internal usage spike after Anthropic's improved Claude Opus 4.6 model launched, and suddenly the crypto exchange's finance team had questions. Hard ones.
The response: a tiered cap system tied to job function. Junior developers get $500 weekly budgets. Senior infrastructure engineers might get $5,000. It's the enterprise equivalent of moving from an open bar to drink tickets.
"Once people understand what's possible, usage takes off on its own. Then the focus shifts from 'Are people using AI?' to 'Are they using it well?'"
This isn't a Coinbase problem. It's an industry reckoning. Deloitte and other major enterprises are implementing similar limits as AI vendors raise prices and companies realize their December 2025 "just try everything" strategy has budget implications. The timing matters: most companies are hitting their first full fiscal year of enterprise AI deployment, which means CFOs are now looking at annualized costs, not pilot budgets.
The deeper signal isn't the price hikes themselves. It's that companies are finally treating AI tools like actual infrastructure with cost-per-query economics. For six months, AI felt free enough to be wasteful. Developers would burn tokens on questions they could Google. Product managers would regenerate the same analysis five times because iterations were instant and seemed costless.
Now comes the forcing function:
- Do you really need Opus 4.6 for that refactoring task, or will a cheaper model work?
- Should that marketing copy generation run through an API call, or can a junior copywriter just write it?
- Which use cases actually drive revenue or efficiency gains worth the compute cost?
These are the right questions. The danger is that companies overcorrect, implementing caps so tight they throttle genuine productivity gains. Witoff's tiered system is smarter than a blanket freeze, but even that assumes organizations can accurately predict which roles need which budgets. In practice, the highest-leverage AI users are often not the most senior, they're the people who figured out how to automate the grunt work everyone else still does manually.
The Implication
Watch for a three-tier market to emerge. Top tier: companies that instrument AI usage well, measure ROI per use case, and allocate budgets based on actual productivity gains. Middle tier: companies that panic, slash budgets uniformly, and lose the high performers who've integrated AI into their workflow. Bottom tier: companies still in denial about AI costs, heading for a nasty Q3 budget meeting.
If you're building AI tooling, the smart play is obvious: give finance teams the visibility they're screaming for. Usage dashboards by department, cost per task type, ROI attribution. The company that makes CFOs feel in control of AI spending will eat the market. The companies that just raise prices and expect enterprises to pay? They're about to learn what happens when procurement gets involved.