Author｜Hualin Wuwang

Editor｜Jingyu

A few days ago, GeekPark reported that Microsoft—having heavily bet on AI—quietly revoked Claude Code licenses for most of its employees.

This move is highly puzzling. After all, in this wave of AI adoption, the biggest marketing pitch to enterprise customers has been “productivity enhancement.” If AI boosts productivity, why would Microsoft stop its employees from using Claude Code?

Microsoft isn’t alone. “Tightening token usage”—discouraging employees from unrestrained “vibe coding”—has become a new trend among Silicon Valley tech giants.

Uber exhausted its entire annual AI token budget in just four months. Salesforce writes Anthropic an annual check worth roughly $300 million. An AI consultant revealed that one of his clients spent as much as $500 million on AI in a single month. Meta even quietly took down its internal “tokenmaxxing leaderboard”—a tool originally designed to encourage employees to use AI more.

Today, enterprises are doing something they wouldn’t have dared imagine just a few years ago:

Restricting—and monitoring—employee AI usage.

So why are big tech firms shifting course?

01 “Tokenmaxxing”: A Mirror of Our Times

To understand today’s cost crisis, we must first clarify what “tokenmaxxing” means.

The term began gaining traction around 2025. Literally, it means “maximizing token usage.” Underlying it is a management logic: Since the company spends heavily on AI tools, employees should use them aggressively—the more you use them, the more “digitally transformed” you appear; the less you use them, the more you’re wasting resources. As a result, many companies introduced usage quotas, leaderboards, and even performance metrics to push employees toward AI adoption.

And what happened?

Employees began using enterprise-grade AI models to check the weather, draft birthday greetings, or decide what to eat for lunch.

A study of 2,444 companies found that for every $1 enterprises spend on AI tokens, $0.44 goes toward fixing bugs generated by AI, $0.27 toward rewriting AI-produced code, and $0.11 toward review and merge delays.

In other words, every dollar spent on AI procurement carries nearly 80% in hidden overhead.

Investor Shruti Gandhi captured this perfectly with an analogy: “Tokenmaxxing enterprises are like companies measuring productivity by how many lights they keep on—spending more doesn’t mean producing more.”

Even more ironically, most of these companies have no idea what their employees are actually doing with AI—or whether those tasks were meaningfully improved by AI at all.

This “spending race” burned from 2024 into 2025—and finally exploded en masse this year. JPMorgan published a blunt, uncomfortably titled report: AI Token Costs Are Eating Into Internet Profits.

Shopify, Spotify, ServiceNow, and Roku all cited AI as a major source of operational expense pressure during recent earnings calls. The industry mood is shifting—from “How cool is our AI?” to “Is this money really worth it?”

02 When CEOs Start Questioning ROI

Only 14% of CFOs say they can see clear, measurable returns on their AI investments.

Uber’s Chief Operating Officer Andrew Macdonald candidly admitted on a podcast that linking individual employee productivity gains to broader business impact remains difficult. “If you can’t demonstrate how AI helped ship valuable features to users, justifying token costs becomes impossible.”

That statement cuts to the core of the enterprise AI dilemma: Individual productivity gains do not equal corporate revenue growth.

An employee may write weekly reports three times faster using AI—but company revenue stays flat. Engineers may generate code twice as fast—but “code churn”—the rate at which AI-generated code gets discarded or rewritten—soars by 800%.

Sophia Velastegui, former Chief AI Officer at Microsoft, said something uncomfortable for many managers: “Most people default to automating tasks they dislike—not the ones most valuable to the company.”

Put plainly, enterprises are automating employees’ “hated work,” not their “revenue-generating work.”

This isn’t a technical issue—it’s a prioritization problem. That’s also why roughly 30% of generative AI projects stall and get scrapped at the proof-of-concept stage—costs are unclear, value is ambiguous, and executives naturally decline renewal.

Salesforce CEO Marc Benioff’s approach is emblematic. Faced with a $300 million annual bill to Anthropic, he expects an “intelligent router”: one that routes queries intelligently—using top-tier models only when warranted, and cheaper, smaller models for routine tasks.

The idea itself isn’t novel—“pay-as-you-go” and “resource optimization” have long been standard practice since the cloud era. But this AI wave hit so fast that companies bought first and thought later—and are now playing catch-up.

03 A Return to Rationality—or the Prelude to Winter?

Microsoft recently canceled most of its enterprise licenses for Claude Code, citing cost concerns as the official reason. This decision sparked significant discussion across the industry—especially given that Microsoft is OpenAI’s largest investor while simultaneously cutting subscriptions to a competitor’s product. How much of this is pure cost discipline, and how much reflects strategic positioning, remains unclear.

Regardless, it signals one thing clearly: Enterprises are voting with their feet.

Harness and CloudZero launched AI cost-management tools almost simultaneously—on May 28—each targeting different aspects: one focuses on real-time AI spending and ROI monitoring; the other introduces an “AI financial control plane” that ties every dollar of AI expenditure directly to concrete business outcomes.

The very emergence of these tools speaks volumes: Market demand exists—and it’s urgent.

HubSpot began overhauling its AI agent pricing model in April, moving away from token-based billing to charging per “resolved conversation” or “generated lead”—a directional shift aligning vendor incentives with buyer outcomes. ServiceNow is making similar adjustments. AI vendors are realizing that if they keep selling “usage” instead of “results,” enterprise customers will inevitably revolt en masse.

Is this adjustment the necessary growing pain of AI industrialization—or the prelude to a deeper crisis?

I lean toward the former. But one detail gives pause: Global AI software spending is projected to reach $2.59 trillion in 2026—a 47% year-on-year increase—yet 94% of engineering leaders report that key ROI metrics remain missing. Spending keeps rising, but no one knows where the money goes—or whether it’s well spent. Unless this contradiction is resolved, the next “tokenmaxxing moment” is merely a matter of time.

A Fortune magazine analysis put it bluntly: “Tokenmaxxing is easy. Redesigning workflows is hard.” Most companies today are optimizing existing processes—not reinventing business models. Yet that’s precisely where AI’s true value lies—and where most enterprises still haven’t arrived.

A return to rationality is welcome. But after rationality returns, enterprises face an even harder question: For our business, should AI be a hammer—or an entirely new framework for thinking?

If we only use AI to do old work faster, the bill will inevitably force us back to confront that question.