Earlier this year, “tokenmaxxing” was the hottest trend in Silicon Valley. The concept was seductively simple: push AI usage as far as it would go. CEOs encouraged — in some cases mandated — employees to integrate AI into every workflow: drafting emails, writing code, summarizing meetings, generating reports, even conducting performance reviews.

The logic seemed sound. If AI makes workers more productive, then more AI equals more productivity. Companies created internal leaderboards ranking employees by AI usage. Managers who resisted were labeled “AI-resistant.” The culture was clear: maximize or fall behind.

Then the bill came due.

Uber reportedly blew through its entire annual AI budget in just a few months. The company had dramatically underestimated how quickly costs would snowball when thousands of employees across engineering, operations, and support all adopted AI tools simultaneously.

They weren’t alone. Several companies quietly scaled back their AI deployments. Some cut Claude and ChatGPT licenses for entire departments. Meta killed its internal AI usage leaderboard — the very tool that had been designed to incentivize tokenmaxxing in the first place.

The root cause is structural. AI models charge per token — small units of text processed during each interaction. A single employee using AI heavily might generate $50-200/month in API costs. Multiply that by 10,000 employees, and the math becomes sobering: $500K to $2M per month in AI expenses alone.

The industry is now pivoting from “use as much as possible” to “use as smartly as possible.” Companies are implementing AI budgets, usage dashboards, and tiered access — treating AI like any other enterprise resource with a cost ceiling.

The honeymoon phase with AI is officially over. What comes next is the harder, less glamorous work of figuring out where AI actually delivers ROI — and where it’s just expensive noise.