Nachdem die großen Silicon Valley-Unternehmen Hunderte von Millionen Dollar an Tokens wild verbrannt haben, beginnen sie nun, das Token-Usage ihrer Mitarbeiter zu beschränken.

Artificial intelligence (AI) automates the work that employees "hate" for companies, not the "money - making work".

A few days ago, GeekPark reported that Microsoft, which heavily relies on AI, secretly cancelled the Claude Code licenses for most of its employees.

This incident is very strange, because in the current wave of AI implementation, the biggest selling point for corporate customers is "efficiency improvement". If AI can improve efficiency, why does Microsoft stop its employees from using Claude Code?

Microsoft is not the only company doing this. "Restricting token consumption" and no longer encouraging employees to do "vibe coding" has now become the new trend among large Silicon Valley companies.

Uber used up its entire annual budget for AI tokens within four months. Salesforce pays about $300 million annually to Anthropic. An AI consultant revealed that one of his clients has monthly AI expenses of $500 million. Meta even secretly took the internal "token - maxxing leaderboard" offline – this list was originally supposed to encourage employees to use AI more.

Now, companies are doing something they wouldn't even dream of a few years ago:

Restrict and monitor employees' use of AI.

Why have large companies suddenly changed their tune?

"Token - maxxing", a reflection of the times

To understand today's cost crisis, one must first understand what "token - maxxing" is.

This term became popular around 2025 and literally means "maximizing token consumption". Behind it is a management logic – since the company has spent a lot of money on AI tools, employees should use them as much as possible. The more one uses, the more "digitally transformed" one is, and those who use less are wasting resources. Many companies have therefore introduced usage quotas, leaderboards, and even performance evaluations to force employees to use AI.

What was the result?

Employees began to use the company's AI models to check the weather, write birthday greetings, or ask what to eat.

A study of 2,444 companies found that for every dollar a company spends on AI tokens, $0.44 is spent on fixing AI - generated errors, $0.27 on rewriting AI - generated code, and $0.11 on review and delay in merging.

This means that behind every dollar of AI purchase costs, there are almost 80% hidden losses.

Investor Shruti Gandhi used an apt metaphor: "Token - maxxing companies are like companies that measure productivity by whether all the lights are on – spending more money doesn't automatically mean more output."

Even more ironic is that most of these companies have no idea what employees are doing with AI, nor do they know if the completion of these tasks by AI has brought any change.

This "money - burning race" lasted from 2024 to 2025 and finally broke out this year. JPMorgan published a strict report with an uncomfortably direct title – "The cost of AI tokens is devouring the profits of the Internet".

Shopify, Spotify, ServiceNow, and Roku mentioned in their quarterly calls that AI is the main cost - pressure factor in their operating expenses. The industry sentiment is shifting from "How cool it is to use AI" to "Is it really worth the money?".

When the CEO doubts the ROI

Only 14% of financial directors say they can see a clear and measurable return on their AI investment.

Andrew Macdonald, the Chief Operating Officer of Uber, said frankly in a podcast that it is difficult for them to link the increase in individual employee productivity with the overall performance of the company. "If you don't see how many valuable functions AI brings to you in user support, it's more difficult to justify the token costs."

This sentence reveals the core problem of corporate AI: An increase in individual efficiency doesn't automatically mean an increase in corporate profits.

Employees write their weekly reports three times as fast with AI, but the company's sales don't change. Engineers generate code twice as fast with AI, but the "scrap rate" of the code – that is, the proportion that is discarded or rewritten – increases by 800%.

Sophia Velastegui, the former head of AI at Microsoft, said something that made many managers uncomfortable: "Most people automatically automate the tasks they don't like, not the ones that are most valuable to the company."

Simply put, companies automate the work that employees "hate", not the "money - making work".

This is not a technical problem but a problem of priorities. That's why about 30% of generative AI projects are already abandoned in the proof - of - concept phase – the costs and value are unclear, and of course, the boss won't renew the license.

The approach of Marc Benioff, the CEO of Salesforce, is representative. Facing an annual bill of $300 million to Anthropic, he expects an "intelligent router": One that can decide which queries are worth processing with a top - tier model and which can be handled with a cheap, small model.

This idea itself is not new – in the cloud - computing era, "usage - based payment" and "resource optimization" were standard practices. But the AI wave came so fast that people bought first and thought later. Now they're just starting to catch up.

Return to rationality or a prelude to a winter?

Microsoft recently canceled most of its corporate licenses for Claude Code. The official reason is cost - effectiveness. This incident has sparked a big discussion in the industry – after all, Microsoft is the biggest investor in OpenAI and is simultaneously cutting subscriptions to competitors. It's hard to say how much of this approach is due to cost considerations and how much is due to strategic considerations.

Anyway, it's a signal: Companies are starting to make rational decisions.

Harness and CloudZero each released an AI cost - management tool almost on the same day – on May 28. One focuses on real - time monitoring of AI expenses and ROI, and the other offers an "AI financial control level" to help companies link every dollar of AI spending with specific business results.

The existence of these two products speaks for itself: The market has a need, and this need is urgent.

HubSpot has adjusted the pricing model for its AI agents since April this year. Instead of calculating by tokens, it now calculates by the "number of resolved conversations" or the "number of generated leads" – this is a groundbreaking change that aligns the seller's interest with the buyer's actual performance. ServiceNow is making similar adjustments. AI providers are starting to understand that corporate customers will protest sooner or later if they continue to sell "usage" instead of "results".

Is this adjustment an inevitable pain in the industrialization of AI or a prelude to a bigger crisis?

I tend to believe the latter. But one detail is concerning: Global spending on AI software is expected to rise to $2.59 trillion in 2026, a 47% increase from the previous year. At the same time, 94% of project managers say that the key ROI indicators are still missing. People are spending more and more money, but no one knows where it's going and if it's worth it – if this problem isn't solved, the next "token - maxxing moment" is just a matter of time.

An analysis in Fortune magazine says directly: "Token - maxxing is easy, re - engineering business processes is difficult." Most companies are currently optimizing existing processes instead of developing new business models. This is the real value of AI, and this is the point that most companies haven't reached yet.

Returning to rationality is good. But after returning to rationality, companies still have to answer a more difficult question: Should AI be more of a hammer or a new framework for our business?

If you only use AI to do the old work faster, one day the bill will inevitably lead you back to this question.

This article is from the WeChat account "GeekPark" (ID: geekpark), written by Hualin Wuwang and edited by Jingyu. Published by 36Kr with permission.