After Burning Billions of Dollars Worth of Tokens, Silicon Valley Giant Now Restricting Employee Token Usage

Original Title: “After Burning Billions of Dollars Worth of Tokens, Silicon Valley Giants Begin Restricting Employee Token Usage”

For enterprises, automation is about tackling the “tedious tasks” of employees, not the “profitable tasks”. A few days ago, it was reported that Microsoft, which had placed a heavy bet on AI, quietly halted most employees’ Claude Code licenses internally.

This was quite bizarre because in the current wave of AI implementation, the biggest selling point to enterprise users is “efficiency”. So, why did Microsoft stop allowing employees to use Claude Code? Microsoft is not alone in this. “Restricting Token Usage” and no longer encouraging employees to engage in extreme Vibe Coding has become the new trend among Silicon Valley giants.

Uber burned through its entire annual AI token budget in four months. Salesforce writes a check to Anthropic for about $300 million each year. An AI consultant revealed that one of his clients had a monthly AI expenditure as high as $500 million. Meta even quietly took down the internal “tokenmaxxing leaderboard” — which was originally intended to encourage employees to use AI more.

Now, companies are doing something that would have been unthinkable a few years ago: Restricting and monitoring employees’ use of AI. Why have the giants all turned in this direction?

01 “Tokenmaxxing,” a Microcosm of the Times

To understand today’s cost crisis, we first need to understand what “tokenmaxxing” is. This term began to gain popularity around 2025, literally meaning “maximizing token usage.” It represents a management logic—since the company spent a lot of money buying AI tools, employees should use them to the fullest. The more you use, the more you prove your “digital transformation”; the less you use, the more you are wasting resources. As a result, many companies established usage quotas, leaderboards, and even performance evaluations to urge employees to use AI.

And the result? Employees have started using the company’s enterprise AI model to check the weather, write birthday wishes, and ask what to eat today. A study of 2444 companies found that for every $1 a company spends on an AI token, $0.44 is used to fix AI-generated bugs, $0.27 is used to rewrite AI-produced code, and $0.11 is spent on review and merge delays. In other words, behind every dollar of AI procurement cost, there is nearly 80% of hidden costs.

Investor Shruti Gandhi used a very accurate analogy: “Tokenmaxxing a company is like measuring productivity by leaving all the lights on—spending more money does not mean producing more.” Most of these companies have no idea what employees are using AI for, let alone whether the tasks are being completed or if any changes are being made because of AI.

This “burning money race” ran from 2024 to 2025 and finally exploded this year. JPMorgan issued a strongly worded report with a blunt title that made people uncomfortable—”AI Token Costs Are Devouring Internet Profits.” Shopify, Spotify, ServiceNow, and Roku all mentioned during their earnings calls that AI has become a major source of operating expense pressure. The industry’s overall sentiment is shifting from “how cool is AI” to “is this money really worth it”.

02 When CEOs Start Questioning ROI

Only 14% of CFOs say they can see a clearly measurable return on investment from AI. Uber’s Chief Operating Officer Andrew Macdonald, in a podcast, said something very candid—they found it challenging to link the improvement in employee personal productivity to the overall business impact of the company. “If you can’t see how AI has helped you push valuable features to users, justifying the token cost becomes even harder.”

This statement highlights the core of the corporate AI dilemma: individual efficiency improvement does not equate to company revenue growth. Employees writing weekly reports using AI have tripled their speed, but the company’s revenue has not changed. Engineers using AI to generate code have doubled their speed, but the “churn rate” of the code—meaning the proportion that is abandoned or rewritten—has increased by 800%.

Microsoft’s former Chief AI Officer, Sophia Velastegui, made a statement that made many managers uncomfortable: “Most people automate tasks they don’t like, rather than the most valuable tasks for the company.” In simple terms, enterprise automation focuses on employees’ “hated work” rather than “money-making work”. This is not a technical issue; it’s a priority issue. It’s also why about 30% of generative AI projects get stuck at the proof of concept stage and are abandoned—costs are unclear, value is unclear, and naturally, bosses don’t renew.

Salesforce CEO Marc Benioff’s approach is quite representative. Facing a $300 million annual Anthropic bill, his expectation is for an “intelligent router”: something that can determine which queries are worth using a top model for and which ones can make do with a cheaper, smaller model. The idea itself is not novel—back in the era of cloud computing, “pay-as-you-go” and “resource optimization” were standard operations. But the AI wave came too fast; everyone bought first and thought later, and now they are just starting to catch up.

03 Rational Regression or Prelude to Winter?

Microsoft recently revoked most of Claude Code’s enterprise licenses, citing cost reasons. This move has sparked considerable discussion in the industry—after all, Microsoft is the largest investor in OpenAI and at the same time is cutting off subscriptions of competing products. It’s hard to say how much of this is cost consideration and how much is strategic positioning.

Regardless, it sends a signal: enterprises are starting to vote with their feet. Almost on the same day, Harness and CloudZero released AI cost management tools, on May 28, each focusing on real-time monitoring of AI expenses and ROI and introducing an “AI financial control plane” to help companies tie every dollar of AI spending to specific business outcomes. The emergence of these two products alone indicates a problem: there is a demand in the market, and it is very urgent.

Starting in April this year, HubSpot adjusted the pricing model for its AI agents, no longer charging per token but instead billing based on “resolved conversation counts” or “generated lead counts”—this is a directional shift that aligns the seller’s interests with the buyer’s actual output. ServiceNow is also making similar adjustments. AI vendors are realizing that if they continue to sell “usage” instead of selling “results,” enterprise customers will eventually push back collectively.

Is this adjustment a necessary growing pain for the industrialization of AI, or the prelude to a larger crisis? I tend to believe it’s the former. However, there is one detail that is slightly concerning: Global AI software spending is expected to reach $2.59 trillion by 2026, a 47% year-over-year increase. Yet, at the same time, 94% of engineering managers indicate that key ROI metrics are still missing. As more money is poured in, but no one knows where it is being spent or if it is worth it—this contradiction, if not resolved, will make the next ‘tokenmaxxing moment’ only a matter of time.

An analysis by Fortune magazine put it bluntly: “Tokenmaxxing is easy, redesigning workflows is hard.” Most companies are currently focused on optimizing existing processes rather than reinventing business models. This is where the true value of AI lies and where most enterprises have not yet reached. A rational return is a good thing. However, after this rational return, companies still need to answer a more challenging question: Should AI for our business be merely a hammer, or should it be a new mindset? If you only use AI to do old work faster, the bill will one day force you to face this question again.

[BlockBeats]