Author｜Hua Lin Dance King

Editor｜Jing Yu

A few days ago, Geek Park reported that Microsoft, which has heavily invested in AI, quietly stopped most employees' access to Claude Code.

This is quite strange, as the biggest marketing point for enterprise users during this wave of AI implementation has been "efficiency improvement". If it can improve efficiency, why did Microsoft stop employees from using Claude Code?

Microsoft is not the only company doing this; "tightening token usage" and no longer encouraging employees to engage in frantic Vibe Coding has become a new trend among major Silicon Valley companies.

Uber spent its entire annual AI token budget in just four months. Salesforce pays about $300 million annually to Anthropic. One AI consultant revealed that one of his clients' monthly AI spending reached $500 million. Meta even quietly took down its internal "tokenmaxxing leaderboard" — that leaderboard was originally meant to encourage employees to use AI more.

Now, companies are doing something they would never have dared to think about a few years ago:

Limiting and monitoring employees' use of AI.

Why are large companies making this shift?

01 "Tokenmaxxing," a microcosm of the era

To understand today's cost crisis, we first need to clarify what "tokenmaxxing" is.

This term started to become popular around 2025, literally meaning "maximizing token usage". Behind it is a management logic—since companies spend a lot of money on AI tools, employees should use them as much as possible. The more you use it, the more you prove you are "digitally transforming"; the less you use, the more resources you waste. As a result, many companies set usage quotas, leaderboards, and even performance evaluations to push employees to use AI.

And the result?

Employees began using the company's enterprise-level AI models to check the weather, write birthday wishes, and ask what to eat today.

A study of 2,444 companies found that for every $1 companies spent on AI tokens, $0.44 went to fixing bugs generated by the AI, $0.27 went to rewriting the code generated by the AI, and $0.11 was consumed in review and merging delays.

In other words, behind every dollar of AI procurement costs, there is almost 80% of implicit loss.

Investor Shruti Gandhi made a very fitting analogy: "Tokenmaxxing companies are like those that measure productivity by having all the lights on—spending more money does not equate to producing more."

More ironically, most of these companies have no idea what tasks employees are using AI for, nor do they know if the completion of those tasks has brought any change due to AI.

This "money-burning competition" has burned from 2024 to 2025 and finally exploded this year. JPMorgan issued a harshly worded report, with a title that is uncomfortable to read—"The Cost of AI Tokens is Eating Up Internet Profits".

Shopify, Spotify, ServiceNow, and Roku all mentioned in their earnings calls that AI has become a major source of pressure on operating expenses. The overall atmosphere in the industry has shifted from "how great AI is" to "is this money really worth spending?".

02 When CEOs start questioning ROI

Only 14% of CFOs say they can see a clear and measurable return on AI investments.

Uber's Chief Operating Officer Andrew Macdonald candidly stated in a podcast that they find it difficult to relate the increase in individual employee productivity to the overall business impact of the company. "If you can't see how AI helps you push valuable features to users, it becomes even harder to justify the token costs."

This statement points out the core of the enterprises' AI dilemma: personal efficiency improvement does not equate to company profit growth.

Employees can write weekly reports three times faster using AI, but the company's revenue remains unchanged. Engineers generate code twice as fast with AI, but the "code attrition rate"—that is, the ratio of code that is abandoned or rewritten—has increased by 800%.

Microsoft's former Chief AI Officer Sophia Velastegui said something that makes many managers uncomfortable: "Most people default to automating tasks they don't like rather than the most valuable tasks for the company."

In other words, businesses are automating employees' "hated tasks," not the "money-making tasks".

This is not a technical issue; it's a prioritization issue. This is also why about 30% of generative AI projects are abandoned once they get stuck in the proof-of-concept stage—costs are unclear and value is also unclear, so naturally, the boss will not renew the subscription.

Salesforce CEO Marc Benioff's approach is quite representative. Faced with an annual $300 million bill from Anthropic, his expectation is for a "smart router" that can determine which queries are worth using top models for and which can be handled by cheaper smaller models.

This idea itself is nothing novel—long before the era of cloud computing, "pay-as-you-go" and "resource optimization" were standard operations. But this wave of AI has come too abruptly, and everyone bought first and thought later; now they're beginning to play catch-up.

03 Rational Return, or Prelude to a Chill?

Recently, Microsoft canceled most enterprise licenses for Claude Code, with official reasons pointing to cost factors. This has sparked significant discussion in the industry—after all, Microsoft itself is the largest investor in OpenAI while simultaneously cutting subscriptions for competitors, making it hard to determine what is cost consideration and what is strategic layout.

But in any case, it represents a signal: companies are voting with their feet.

Harness and CloudZero both released AI cost management tools almost on the same day—May 28—one focusing on real-time monitoring of AI spending and ROI, the other offering an "AI Financial Control Plane" to help companies link every dollar of AI expenditure to specific business outcomes.

The emergence of these two products indicates a problem: there is a demand in the market, and the demand is urgent.

HubSpot started adjusting its pricing model for AI agents from April this year, no longer charging by token but by the "number of conversations solved" or "number of leads generated"—this is a directional shift aligning the interests of sellers with the actual outputs of buyers. ServiceNow is also making similar adjustments. AI vendors are realizing that if they continue to sell "usage" rather than "results," enterprise clients will eventually push back collectively.

Is this adjustment a necessary pain for the industrialization of AI, or a prelude to a greater crisis?

I tend to think it is the former. However, one detail raises some concerns: global AI software spending is expected to reach $2.59 trillion by 2026, a 47% increase year-on-year, but at the same time, 94% of engineering heads report a lack of key ROI metrics. The more money is spent, the more unclear it is where it is being burned and whether it is worth it—if this contradiction is not resolved, the next "tokenmaxxing moment" is only a matter of time.

An analysis from Fortune magazine puts it very directly: "Tokenmaxxing is easy; redesigning workflows is hard." Most companies are currently optimizing existing processes rather than reinventing business models. This is where the true value of AI lies, and it is also where most enterprises have yet to arrive.

A rational return is a good thing. But after the rational return, companies still need to answer a more difficult question: Should AI serve as a hammer for our business, or as a new framework of thinking?

If it is only about using AI to do old tasks faster, one day the bills will force you back to this question.