"When Token is more expensive than people, "AI narrative" runs into trouble."

The rationality of enterprise AI spending is facing a severe test, with Token consumption continuously increasing, but the quantifiable business value is hard to find. On May 22, Andrew Macdonald, Chief Operating Officer of Uber with a market value of over 200 billion US dollars, publicly stated in a podcast that there is still no correlation between the growth of Token consumption and actual product improvement. Macdonald pointed out that it is increasingly difficult for companies to justify the continuously rising AI spending. He even coined a term "tokenmaxxing" for wasteful practices within the engineering team. Earlier in mid-May, Microsoft began cutting internal Claude Code authorizations citing "difficulty in sustainability" due to Token bills. These two events combined have forced the market to confront a variable that was previously overlooked. Token economics, which refers to the unit economics of Token consumption at the enterprise scale, has now become a core pillar of the entire AI investment discourse. Five sets of data paint a new picture Since April, multiple sets of data have emerged, collectively portraying a worrisome scenario. In April this year, Uber's Chief Technology Officer publicly stated that the company had burned through its entire yearly Claude Code budget in just four months. Among 5,000 engineers, the monthly utilization rate ranged from 84% to 95%, with individual monthly bills ranging from $150 to $2,000. The CTO himself reportedly consumed tokens worth $1,200 during a two-hour internal presentation. Macdonald described being "shocked and speechless" upon learning this number. On the Microsoft side, according to a newsletter by Tom Warren of The Verge, Claude Code quickly gained popularity among internal engineers, but the consumption model based on Tokens made sustained spending unsustainable, leading Microsoft to begin cutting related authorizations. GitHub announced that starting from June 1, all Copilot plans would switch from fixed subscriptions to usage-based billing. An official discussion thread received nearly 900 opposing votes, with users calculating that a single AI programming session typically costs $30 to $40, meaning a $10 monthly plan would be depleted in one use. The developer productivity platform Entelligence.AI, after aggregating data from 2,444 enterprises, found that: For every $1 spent on AI Token fees, only 18 cents generated actual value for reaching users. 44 cents were used for fixing bugs introduced by AI; 27 cents went towards rework; and 11 cents were spent on review friction. According to the Bloomberg Silicon Data LLM Token Expenditure Index, Token prices have risen by about 65% since the end of February, with US AI software prices seeing an accumulated increase of 20% to 37% over the past year. Bull vs. bear: Two interpretations of the same fact The same data points to vastly different conclusions under different analytical frameworks. Bullish views believe that the current upheaval is simply a painful period of successful transformation. According to Jim Schneider of Goldman Sachs in early May, by 2030, agent-based AI will drive a 24-fold increase in Token consumption, reaching around 1.2 quadrillion Tokens per month, leading to positive gross margins for super-scale cloud service providers and model providers in the next 3 to 12 months. Rich Privorotsky of Goldman Sachs believes that by the first quarter of 2026, "Token maximization" as a key performance indicator may have peaked, and the industry is shifting from pursuing consumption quantity to a healthier metric of "unit effective action cost." Research by JPMorgan also found a significant spike in the addition and updating of Python packages on PyPI in early 2026, a trend that was not seen when ChatGPT launched in 2022, indicating real productivity improvements are taking place. Furthermore, Mag 7 currently has a price-to-earnings ratio of around 20 times forward earnings, much lower than the peak of the 2000 tech bubble at 52 times, Japan in 1989 at 67 times, and the "Nifty Fifty" era at 34 times. By historical bubble standards, the current situation does not constitute a bubble. Bearish views, as articulated most comprehensively by Jim Covello, a semiconductor analyst at Goldman Sachs, in an April report. He pointed out that almost all the value in the AI supply chain flows to semiconductor companies, a phenomenon unprecedented in history and unsustainable. Chip companies should benefit when customers benefit, but in this cycle, their prosperity comes at the expense of consumption in the entire upstream industry chain. NVIDIA's net profit has increased by about 20 times since the launch of ChatGPT; major super-scale cloud service providers have exhausted their operating cash flow and turned to debt - data center-related debt issuances in 2025 amounted to about $182 billion, doubling from 2024. MIT's Nanda research shows that 95% of enterprises investing in generative AI have zero returns. This decoupling may persist for a while, but cannot last indefinitely. Concerns about the circular financing structure This discussion also involves a more complex dimension: the financial loop between super-scale cloud service providers and AI labs. According to corporate disclosure documents compiled by The Information, OpenAI and Anthropic account for over half of the future cloud service commitments of Microsoft, Oracle, Google, and Amazon, totaling around $2 trillion. Specifically: Out of Microsoft's $627 billion backlog of cloud services, $280 billion is tied to OpenAI; Out of Oracle's $553 billion pipeline business, 54% (about $300 billion) is committed by OpenAI; Out of Google's $467.6 billion, Anthropic accounts for 43% (about $200 billion); Amazon's corresponding exposure reaches 51% of its $464 billion backlog. This financing structure has inherent cyclicality. Microsoft's $13 billion investment in OpenAI is mainly redeemed in Azure credits, which OpenAI uses to purchase Azure computing power, subsequently included in cloud revenue by Microsoft. Similarly, super-scale cloud service providers act as both equity investors in AI labs and service suppliers charging for computing power. This structure is also reflected in profit data. Alphabet reported record profits of $62.6 billion in the first quarter, with about $28.7 billion, nearly half coming from unrealized gains on the shares held by Anthropic. Out of Amazon's $30.3 billion first-quarter profit, $16.8 billion is pre-tax unrealized income from Anthropic, while its free cash flow plummeted by 95% to $1.2 billion from the previous quarter, with data center capital expenditures reaching $44.2 billion. The sustainability of this system depends on AI labs' continuous ability to obtain external financing to fulfill cloud computing commitments, which in turn relies on enterprise customers' willingness to pay increasing Token bills. Reportedly, Anthropic's cost is as high as $3 for every $1 of revenue. Once the pace of financing slows down, the credibility of cloud revenue predictions will decrease, and the valuation multiples of super-scale cloud giants will come under pressure for reassessment. The chain of this structure is bidirectional, and will also face a bidirectional rupture. This is not 1999, but the problem is real The current situation does not constitute a typical bubble setup. In terms of valuation multiples, the seven tech giants currently correspond to approximately 20 times forward earnings, much lower than the peaks of the 2000 tech bubble at 52 times, the Japanese market in 1989 at 67 times, and the "Nifty Fifty" era at 34 times. AI technology itself is real. For heavy users, the data on productivity improvements is verifiable. OpenAI has an annualized revenue of around $20 billion, while Anthropic has about $4.3 billion, and both labs are not disappearing. Now, the cost of Tokens (computing expenses) has become a key factor in determining the success of AI, whereas six months ago, people barely talked about this topic. At that time, everyone only cared about whether "the technology worked." Now the answer is clear: for specific tasks and specific user groups, the technology does work. But a new problem has emerged: can the money that enterprises save using AI trickle up in time to outperform the valuation gap left by the capital markets for AI labs and cloud giants? Optimists believe that as long as the technology continues to mature, enterprise ROI (return on investment) will turn positive within 1 to 1.5 years. Pessimists believe that more executives will openly complain about the low ROI of AI investments like MacDonald and begin cutting budgets. Both possibilities are happening, and the outcome is uncertain. The only thing that is certain is that the lie that "as long as Token consumption is increasing, the AI transformation is successful" is shattered. Token consumption does not equate to business value, and these two bubbles will eventually burst. The AI bill has come due, but who will ultimately foot that bill? It is still an unknown variable. This article is a translation of a piece originally published on Wall Street View, GMTEight editor: Chen Wenfang.