--- title: "Jensen Huang's Token Economics Falters! Microsoft and Amazon Jump Ship" type: "News" locale: "en" url: "https://longbridge.com/en/news/288262475.md" description: "Multiple tech giants are adjusting their strategies as AI usage costs spiral out of control. Amazon employees frantically burned through tokens to climb internal leaderboards, causing bills to surge; meanwhile, Microsoft canceled most of its Claude licenses, requiring a migration to its own products. This reflects a shift in corporate sentiment from worrying about underutilization of AI to guarding against the high costs of overuse, prompting a reevaluation of the return on investment for AI initiatives" datetime: "2026-06-01T11:14:06.000Z" locales: - [zh-CN](https://longbridge.com/zh-CN/news/288262475.md) - [en](https://longbridge.com/en/news/288262475.md) - [zh-HK](https://longbridge.com/zh-HK/news/288262475.md) --- # Jensen Huang's Token Economics Falters! Microsoft and Amazon Jump Ship Token economics may not be so economical after all. (Except for Jensen Huang and his AAAAA-grade GPUs) Recently, Axios reported a shocking incident: one company racked up a $500 million bill for Claude in just one month, equivalent to RMB 3.4 billion. The reason? After the boss enabled enterprise licenses for Anthropic for employees, they forgot to set usage caps, and tokens kept burning... After one month, RMB 340 million vanished into thin air. Setting aside whether it was worth it, if we calculate based on Anthropic's current annualized recurring revenue (ARR) of approximately $47 billion, this single company contributed nearly **one-eighth of Anthropic's monthly revenue** in just one month. In other words, for every $8 Anthropic earns, $1 comes from this company. Even more astonishingly, the identity of this company has not yet been disclosed. However, Axios noted in its report that there are actually very few companies globally that can "painlessly" absorb a $500 million AI bill. Soon, various speculations began appearing on X, with Amazon being the most frequently named suspect. Coincidentally, around the same time, the Financial Times revealed that Amazon had canceled its internal AI usage leaderboard because employees were frantically burning through tokens to climb the ranks, even executing numerous tasks with no actual value. Viewing these two pieces of news together changes the narrative significantly. Over the past two years, companies' biggest concern was that employees were not using AI. Now, more and more companies are starting to worry about another issue: Is AI being used too much? ## **US Tech Giants Start Crunching the Numbers** This shift is occurring with increasing frequency. Microsoft is a typical example. Recently, **Microsoft** announced that it would **cancel** most Claude Code licenses for its Experiences + Devices division (responsible for Windows, Microsoft 365, Outlook, Teams, and Surface) by June 30, requiring engineers to **migrate** to its own GitHub Copilot CLI. It has only been six months since Claude Code was introduced internally at Microsoft. Microsoft's reasoning aligns with traditional Silicon Valley culture— Claude Code has completed its phase of helping teams learn and explore; it is now time to return to "Eat Your Own Dog Food" and use their own products. Regardless, token bills remain an unavoidable topic. Similar changes are also evident at **DeepSeek**. When releasing V4 in April this year, DeepSeek mentioned in its technical report that V4 had become the Agentic Coding model used daily by internal employees, offering a better experience than Claude Sonnet 4.5 and delivery quality close to Opus 4.6. Although the official statement did not explicitly mention cost factors, for a company with self-developed models, using its own models to accomplish the same tasks is obviously a more economical choice. If Microsoft and DeepSeek were relatively subtle and polite, Uber was more direct. Uber CTO Praveen Neppalli Naga revealed earlier this year that company engineers burned through the entire annual Claude Code budget in just four months. Subsequently, Uber COO Andrew Macdonald publicly stated: **There seems to be no obvious linear relationship between AI token consumption and the final release of valuable products.** In other words, spending more tokens does not necessarily mean creating more value. Similar reflections are beginning to emerge within more companies. Previously, **Duolingo** had planned to include AI usage in employee performance evaluations. However, after employees questioned whether they were "forced to use AI for the sake of using AI," the company ultimately withdrew this decision. "It felt like we weren't being held accountable for actual results, but rather trying to push something that isn't always suitable." Louis von Ahn, CEO of Duolingo, summarized it this way in a podcast in April this year. **Meta**'s changes are even more representative. After it was exposed that Meta had established an internal Claude usage leaderboard, consuming billions of tokens in a single month, Meta gradually began tightening related incentives, shifting from encouraging "more usage, more grinding" to focusing more on actual output. Meanwhile, similar voices have emerged domestically. Zheng Yinhe from **miHoYo** once shared an experience: after launching an Agent project, it burned through RMB 2 million in token fees in a single night. This tuition fee was not cheap. But it made more and more companies realize a problem: Tokens themselves are not value; completing tasks, delivering products, and generating revenue are what constitute value. It can be said that companies still believe in AI, but compared to last year, they are no longer simply pursuing higher token consumption. Instead, they are seriously calculating the ROI behind every token. ## **AI Has Become a Financial Issue for the First Time** Undoubtedly, after more than half a year of "Huang-style Token Economics," US tech giants are beginning to seriously reflect: Are these tokens really worth the spend? In March this year, Jensen Huang publicly endorsed this logic on the "All-In Podcast." If an engineer with an annual salary of $500,000 consumes less than $250,000 worth of tokens annually, he would be deeply concerned. At the time, this statement was regarded as gospel by many enterprises. After all, if AI truly improves engineer efficiency, then burning more tokens is essentially purchasing productivity. The more you use, the more you save—that's the logic. But the problem is: when real bills start arriving at companies, things become not so simple. Whether it is Amazon canceling its leaderboard, Microsoft shrinking its Claude Code licenses, or Uber discovering that engineers burned through the annual budget in four months. Bosses ultimately voted with their feet. Recently, discussions on this matter have also begun to ferment on Hacker News. Some believe this is an important turning point. The frenzied phase of equating token consumption with AI adoption rates, or even with productivity, is coming to an end. Others have directly pointed the finger at the "Tokenmaxxing" culture that has been popular over the past six months. _(Note: A token is the basic unit for large language models to process text and also the billing unit. Tokenmaxxing refers to companies and employees frantically pursuing token consumption, treating "how many tokens were burned" as a measure of AI adoption rate and productivity.)_ In their view, the problem lies not with AI, but with companies mistakenly treating "burning more tokens" as the goal itself. If more efficient models were used, agent workflows were reasonably controlled, or humans were involved in key decisions, costs could have been much lower. Of course, some joked: The biggest winner of this movement, from start to finish, might only be Jensen Huang. However, another viewpoint is equally worth noting. Many developers believe this is not a signal that the AI hype is fading. Quite the opposite. It means companies are finally moving from the "use first, talk later" phase into a stage of "refined operations." In the future, the focus may no longer be on who burns more tokens, but on who can complete more tasks with fewer tokens. For example: more efficient agent workflows; using cheaper models for simple tasks; reserving expensive models only for key decisions; and stricter budget and permission management. Risk Warning and Disclaimer The market involves risks, and investment requires caution. This article does not constitute personal investment advice, nor does it take into account the specific investment objectives, financial status, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article align with their specific circumstances. Investment decisions made based on this content are the sole responsibility of the investor. ### Related Stocks - [MSFT.US](https://longbridge.com/en/quote/MSFT.US.md) - [AMZN.US](https://longbridge.com/en/quote/AMZN.US.md) - [NVDA.US](https://longbridge.com/en/quote/NVDA.US.md) - [UBER.US](https://longbridge.com/en/quote/UBER.US.md) - [DUOL.US](https://longbridge.com/en/quote/DUOL.US.md) - [META.US](https://longbridge.com/en/quote/META.US.md) - [NVD.DE](https://longbridge.com/en/quote/NVD.DE.md) ## Related News & Research - [Was Amazon's Tokenmaxxing Fiasco Behind Claude's $500M Mystery Bill?](https://longbridge.com/en/news/288103609.md) - [Dear Microsoft Stock Fans, Mark Your Calendars for June 2](https://longbridge.com/en/news/288313881.md) - [Microsoft and Google are late to AI coding, but 'absolutely critical' they compete for growth](https://longbridge.com/en/news/288322876.md) - [Alipay launches AI payment solution, AI wallet, Token Pay services as AI commerce expands](https://longbridge.com/en/news/287599126.md) - [Does Netskope’s Claude Integration Shift Its AI Governance Edge Into a Lasting Advantage for NTSK?](https://longbridge.com/en/news/288154220.md)