When tokens are more expensive than people, the "AI narrative" runs into trouble.
The rationality of corporate AI spending is being severely tested. Token consumption continues to climb, but quantifiable commercial value remains elusive.
On May 22, Andrew Macdonald, COO of Uber, which has a market cap of more than $200 billion, publicly stated on a podcast that the correlation between the growth of token consumption and substantial product improvements "simply does not exist."
Macdonald noted that it is increasingly difficult for the company to justify the rising AI expenditure. He even coined a term for the engineering team's internal waste: "tokenmaxxing."
Earlier in mid-May, Microsoft began cutting internal Claude Code licenses, citing "unsustainable" token billing.
The convergence of these two events has forced the market to confront a previously overlooked variable. Token economics—specifically, the unit economics of token consumption at the enterprise scale—has shifted from a peripheral topic to the core pillar supporting the entire thesis of AI investment.
Five Data Sets Paint a New Picture
Since April, multiple sets of data have emerged, together outlining an alarming landscape.
In April this year, Uber’s CTO publicly stated the company burned through its annual Claude Code budget in just four months.
Among 5,000 engineers, monthly usage rates ranged from 84% to 95%, with individual monthly bills varying from $150 to $2,000. The CTO reportedly consumed $1,200 worth of tokens during a single two-hour internal demo.
Macdonald described himself as "literally speechless" upon learning these numbers.
As for Microsoft, according to Tom Warren’s Notepad newsletter from The Verge, Claude Code became popular among internal engineers, but the token-based billing model made large-scale spending unsustainable, prompting Microsoft to cut related licenses.
GitHub has announced that from June 1 all Copilot plans will switch from fixed subscriptions to usage-based billing.
An official discussion thread garnered nearly 900 downvotes, as some users calculated a single AI programming session typically consumes $30 to $40, meaning a $10 monthly subscription would be exhausted in one use.
Developer productivity platform Entelligence.AI analyzed data from 2,444 companies and found:
For every $1 spent on AI token fees, only 18 cents generated actual user-facing value.44 cents went to fixing bugs introduced by AI; 27 cents to rework; 11 cents consumed by review friction.
According to Bloomberg Silicon Data LLM Token Expenditure Index, token prices have risen about 65% since the end of February this year, with US AI software prices up 20% to 37% over the past year.
Bull-Bear Debate: Two Interpretations of the Same Facts
The same data leads to completely different conclusions under differing analysis frameworks.
Bulls argue the current chaos is simply transition pains for a successful transformation.
Goldman Sachs’ Jim Schneider assessed in early May that by 2030, agent-style AI will drive token consumption to grow 24-fold, reaching about 120 zettatokens per month, and gross margins for major cloud and model providers will turn positive in the next 3 to 12 months.
Goldman Sachs’ Rich Privorotsky suggests that Q1 2026 may be the peak of "tokenmaxxing" as a KPI, and the industry is shifting from maximizing consumption to measuring "unit effective action cost," a healthier metric.
Economic research from JP Morgan also shows a spike in new and updated Python packages on PyPI in early 2026, a trend not seen after the launch of ChatGPT in 2022, indicating legitimate productivity gains are happening.
Additionally, the current price-to-earnings ratio of the Mag 7 is about 20x forward earnings, far below the peak of 52x during the 2000 tech bubble, Japan’s 67x in 1989, and the "Nifty Fifty" era’s 34x. By historical bubble standards, the current market does not constitute a bubble.
Bears, led by Goldman Sachs semiconductor analyst Jim Covello in an April report, present the most systematic view.
He points out that almost all the value in the AI supply chain has flowed to semiconductor companies, an unprecedented and unsustainable phenomenon. Chipmakers should profit when their customers benefit, but in this cycle, their prosperity comes at the expense of upstream consumption along the whole supply chain.
Nvidia’s net profits have grown around 20-fold since ChatGPT launched; major hyperscale cloud providers have burned through operating cash flow and resorted to borrowing—data center-related debt issuance is about $182 billion in 2025, doubling from 2024.
MIT Nanda research shows that 95% of companies investing in generative AI have seen zero returns. This decoupling may last for a while, but can't continue forever.
Risks of Circular Financing Structure
The discussion also concerns a more complex layer: the financial loop between hyperscale cloud providers and AI labs.
According to corporate disclosures compiled by The Information, OpenAI and Anthropic together account for over half of $2 trillion in future cloud commitments from Microsoft, Oracle, Google, and Amazon. Specifically:
Microsoft: Of $627 billion in cloud backlog orders, $280 billion are tied to OpenAI;Oracle: Of $553 billion in pipeline business, 54% (around $300 billion) is committed by OpenAI;Google: Of $467.6 billion, Anthropic accounts for 43% (about $200 billion);Amazon: Of the $464 billion backlog, 51% is similarly exposed.
This financing structure is self-reinforcing. Microsoft’s $13 billion investment in OpenAI is mainly fulfilled in Azure credits, which OpenAI uses to buy Azure compute, and Microsoft then records as cloud revenue.
The same hyperscale cloud providers are both equity investors in AI labs and suppliers collecting compute bills.
This structure is also reflected in profit data. Alphabet reported a record $62.6 billion first-quarter profit, of which about $28.7 billion—nearly half—came from the mark-to-market gain in Anthropic holdings.
Amazon’s $30.3 billion first-quarter profit included $16.8 billion in pre-tax unrealized gains from Anthropic, while free cash flow plummeted 95% to $1.2 billion due to $44.2 billion in data center capital expenditure.
The sustainability of this structure depends on the continued ability of AI labs to secure external financing to fulfill their cloud computing commitments, which in turn relies on enterprise clients’ willingness to pay rising token bills.
Reportedly, Anthropic currently spends $3 in costs for every $1 of revenue. If the funding pace slows, the credibility of cloud revenue forecasts will fall, and the valuation multiples of hyperscale cloud providers will face re-rating pressure.
This chain can break in either direction.
This Is Not 1999, but Real Problems Exist
The current situation does not constitute a classic bubble.
From a valuation perspective, the Tech Seven now trade at about 20x forward earnings, well below the peaks during the 2000 tech bubble (52x), 1989 Japan (67x), or the "Nifty Fifty" (34x).
AI technology itself is real. For heavy user groups, productivity gains are verifiable. OpenAI’s annualized revenue is about $20 billion, Anthropic’s about $4.3 billion; neither lab is disappearing anytime soon.
Today, token costs (compute expenditure) have become the key to determining AI's success or failure, and half a year ago, this topic was barely discussed.
Back then, everyone only cared about whether "the technology works." Now the answer is clear: for specific jobs and specific people, the technology really works.
But a new question arises: Can the savings achieved by downstream enterprises using AI be passed upstream quickly enough to beat the valuation window the capital markets have given to AI labs and cloud giants?
Optimists believe that as technology matures, enterprise ROI can turn positive within 1 to 1.5 years.
Pessimists think more execs, like Macdonald, will publicly complain about AI’s poor input/output ratio and start slashing budgets.
Both outcomes are happening, and the result is still undecided. The only certainty is that the old myth “as long as token consumption is rising, AI transformation is a success” has been shattered.
High token consumption does not equal commercial value—these two bubbles must ultimately be squeezed out. The AI bill is due, but who's going to pay it? The answer is still unknown.
Risk Warning and DisclaimerThe market carries risks; investments should be made cautiously. This article does not constitute personal investment advice and does not consider the special investment objectives, financial circumstances, or needs of individual users. Users should consider whether any opinions, viewpoints, or conclusions contained herein suit their situation. If you invest based on the above, you do so at your own risk.