Author: Bao Yilong
Source: Wall Street News
The justification for corporate AI spending is facing a severe test, as Token consumption continues to climb, yet quantifiable commercial value remains elusive.
On May 22, Uber's Chief Operating Officer Andrew Macdonald, whose company is valued at over $200 billion, stated publicly on a podcast that the link between the growth in token consumption and substantial product improvement "doesn't exist yet."
Macdonald pointed out that companies are finding it increasingly difficult to rationalize the continuously rising AI expenditures. He even coined a term for the wasteful phenomenon within engineering teams: "tokenmaxxing."
Earlier in mid-May, Microsoft began cutting internal Claude Code licenses, citing token bills as "unsustainable."
The combination of these two events forces the market to confront a previously overlooked variable. Token economics, specifically the unit economics of token consumption at enterprise scale, has evolved from a peripheral issue to the central load-bearing pillar of the entire AI investment thesis.
Five Data Points, Painting a New Picture
Since April, multiple data points have emerged successively, collectively sketching an alarming picture.
In April this year, Uber's Chief Technology Officer publicly stated that the company had 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 himself reportedly consumed $1,200 worth of tokens during a two-hour internal demonstration.
Macdonald described being "speechless" upon hearing this number.
Regarding Microsoft, according to a report in The Verge's Tom Warren's Notepad newsletter, Claude Code quickly became popular among Microsoft's internal engineering teams. However, the token-based billing model made scaled spending unsustainable, prompting Microsoft to proceed with cutting related licenses.
GitHub announced that starting June 1, all Copilot plans would shift from a fixed subscription model to usage-based billing.
The official discussion thread garnered nearly 900 downvotes, as users calculated that a single AI programming session typically consumes $30 to $40, meaning a $10 monthly subscription could be exhausted in a single use.
Developer productivity platform Entelligence.AI aggregated data from 2,444 companies and found:
- For every $1 spent on AI token costs, only 18 cents generated actual value reaching users.
- 44 cents were used to fix bugs introduced by the AI itself; 27 cents went to rework; 11 cents were consumed by review friction.
According to Bloomberg's Silicon Data LLM Token Expenditure Index, token prices have risen about 65% since the end of February this year, and US AI software prices have increased by 20% to 37% cumulatively over the past year.
Bull vs. Bear Debate: One Fact, Two Interpretations
The same data points to starkly different conclusions under different analytical frameworks.
The bullish view argues that the current chaos is merely the growing pains of a successful transformation.
According to Goldman Sachs' Jim Schneider in early May, by 2030, agentic AI will drive a 24-fold increase in token consumption, reaching approximately 120 sextillion tokens per month. The gross margins of hyperscale cloud providers and model vendors will turn positive within the next 3 to 12 months.
Goldman's Rich Privorotsky believes that Q1 2026 might have been the peak for "token maximization" as a KPI. The industry is shifting from pursuing consumption volume to the healthier metric of "cost per effective action."
JP Morgan's economic research also found a jump in new and updated Python packages on PyPI in early 2026, a trend not seen when ChatGPT launched in 2022, indicating that real productivity gains are occurring.
Furthermore, the Magnificent 7 currently trades at about 20 times forward earnings, far below the 52 times at the peak of the 2000 tech bubble, 67 times for Japan in 1989, and 34 times during the "Nifty Fifty" era. By historical bubble standards, this does not constitute a bubble.
The bearish view was most systematically articulated by Goldman Sachs semiconductor analyst Jim Covello in an April report.
He pointed out that almost all value in the AI supply chain flows to semiconductor companies, a phenomenon unprecedented and unsustainable in history. Chip companies should benefit when their customers benefit, but in this cycle, their prosperity comes at the expense of consumption across the entire upstream industry chain.
Nvidia's net profit has grown about 20-fold since ChatGPT's launch; major hyperscale cloud providers have burned through their operating cash flow and are turning to debt—data center-related debt issuance in 2025 was approximately $182 billion, doubling from 2024.
MIT Nanda research shows 95% of enterprises investing in generative AI see zero return. This decoupling may persist for a while, but cannot last forever.
Concerns of the Circular Financing Structure
This discussion touches on a more complex level: the financial loop between hyperscale cloud providers and AI labs.
According to corporate disclosure documents compiled by The Information, OpenAI and Anthropic account for more than half of the approximately $2 trillion in future cloud service commitments from Microsoft, Oracle, Google, and Amazon. Specifically:
- Of Microsoft's $627 billion cloud service backlog, $280 billion is tied to OpenAI;
- Of Oracle's $553 billion pipeline business, 54% (approx. $300 billion) is committed by OpenAI;
- Of Google's $467.6 billion, Anthropic accounts for 43% (approx. $200 billion);
- Amazon's corresponding exposure also reaches 51% of its $464 billion backlog.
This financing structure is inherently circular. Microsoft's $13 billion investment in OpenAI was largely delivered in the form of Azure credits, which OpenAI used to purchase Azure compute. Microsoft then booked this as cloud revenue.
The same hyperscale cloud providers are both equity investors in the AI labs and service providers collecting compute bills.
This structure is also reflected in profit data. Alphabet reported a record Q1 profit of $62.6 billion, of which about $28.7 billion, nearly half, came from the paper appreciation of its Anthropic stake.
Amazon's Q1 profit of $30.3 billion included $16.8 billion in pre-tax unrealized gains from Anthropic, while its free cash flow plummeted 95% to $1.2 billion due to data center capital expenditures of $44.2 billion in the same period.
The sustainability of this system depends on AI labs' continued ability to secure external financing to fulfill cloud computing commitments, which in turn relies on enterprise customers' continued willingness to pay rising token bills.
It is reported that Anthropic currently incurs costs of $3 for every $1 of revenue. Once the pace of financing slows, the credibility of cloud revenue projections will decline, and the valuation multiples of hyperscale cloud vendors will also face re-evaluation pressure.
This chain transmits in both directions and will break in both directions.
This Isn't 1999, But the Problem is Real
The current situation does not constitute a typical bubble setup.
From a valuation multiple perspective, the Tech 7 currently trades at about 20 times forward price-to-earnings, far below the 52 times at the peak of the 2000 tech bubble, 67 times for the Japanese market in 1989, or the 34 times during the "Nifty Fifty" era.
AI technology itself is real. For heavy user groups, data on productivity gains is verifiable. OpenAI has an annualized revenue of about $20 billion, Anthropic about $4.3 billion; these two labs are not going to disappear.
Today, token cost (compute expense) has become the key determinant of AI success or failure. Six months ago, people weren't even discussing this topic.
Back then, people only cared about "whether the technology works." Now the answer is clear: in the eyes of specific jobs and specific people, the technology indeed works.
But a new question arises: Can the money saved by downstream companies using AI be transmitted upward in time to outrun the valuation window the capital market has left for AI labs and cloud giants?
Those bullish on AI believe that as long as the technology continues to mature, corporate ROI (Return on Investment) will turn positive within 1 to 1.5 years.
The bearish believe more executives will follow Macdonald's lead, publicly complaining about low AI ROI and starting to cut budgets.
Both scenarios are playing out; the outcome is undecided. The only certainty is that the old lie—"as long as token consumption is rising, it means the AI transformation is successful"—has been shattered.
High token consumption does not equal commercial value; this bubble must eventually be squeezed out. The bill for AI has come due, but who will ultimately pay for it? That remains an unknown for now.
