Uber and Microsoft Highlight Rising AI Token Costs and Diminishing Returns

Article by Bao Yilong

Source: Wall Street Journal

The justification for enterprise AI spending is under severe scrutiny, as token consumption continues to rise, yet measurable business value remains elusive.

On May 22, Andrew Macdonald, Chief Operating Officer of Uber, which has a market capitalization of over $200 billion, stated on a podcast that "there is no line yet" between the growth in token consumption and meaningful product improvements.

MacDonald noted that companies are finding it increasingly difficult to justify their continuously rising AI expenses. He even coined a term specifically for waste within engineering teams: "tokenmaxxing."

In mid-May, Microsoft began reducing internal Claude Code licenses, citing unsustainable Token billing.

The combination of these two events has forced the market to confront a previously overlooked variable: token economics—the unit economics of token consumption at scale—has risen from a marginal topic to the central pillar of the entire AI investment thesis.

Since April, a series of data sets have been released, collectively painting a concerning picture.

In April this year, Uber's Chief Technology Officer publicly stated that the company had burned through its entire annual Claude Code budget in four months.

Among 5,000 engineers, monthly usage rates ranged from 84% to 95%, with per-person monthly bills ranging from $150 to $2,000; the CTO himself reportedly consumed $1,200 worth of tokens during a two-hour internal demo.

Macdonald described being utterly speechless upon learning this number.

Regarding Microsoft, according to Tom Warren’s Notepad newsletter from The Verge, Claude Code has rapidly gained popularity among Microsoft’s internal engineers, but the token-based billing model has made large-scale spending unsustainable, prompting Microsoft to move toward reducing related licenses.

GitHub has announced that, effective June 1, all Copilot plans will transition from fixed subscriptions to pay-as-you-go billing.

The official discussion post received nearly 900 downvotes because users calculated that a single agent programming session typically consumes $30 to $40, meaning a $10-per-month plan is exhausted in a single use.

After aggregating data from 2,444 companies, the developer productivity platform Entelligence.AI found:

• For every dollar spent on AI Token fees, only 18 cents generated actual value for users.
• 44 cents are spent fixing bugs introduced by AI itself; 27 cents go toward rework; 11 cents are consumed by review friction.

According to Bloomberg’s Silicon Data LLM Token Expenditure Index, token prices have risen approximately 65% since the end of February this year, while U.S. AI software prices have increased cumulatively by 20% to 37% over the past year.

The same data can lead to entirely different conclusions under different analytical frameworks.

Bullish perspectives view the current turmoil as a temporary pain during a successful transition.

According to Jim Schneider of Goldman Sachs in early May, by 2030, agentive AI will drive token consumption growth of 24 times, reaching approximately 120 sextillion tokens per month, with supercloud providers and model providers achieving positive gross margins within the next 3 to 12 months.

Goldman Sachs’ Rich Privorotsky believes that the first quarter of 2026 may already mark the peak of "token maximization" as a KPI, as the industry shifts toward the healthier metric of "cost per effective action."

JPMorgan Chase's economic research also found a sharp increase in new and updated Python packages on PyPI in early 2026, a trend that did not occur when ChatGPT launched in 2022, indicating that genuine productivity gains are taking place.

Moreover, the current P/E ratio of the Mag 7 is approximately 20 times forward earnings, significantly lower than the 52 times seen at the peak of the 2000 tech bubble, 67 times during Japan’s 1989 bubble, and 34 times during the "Nifty Fifty" era. By historical bubble standards, the current valuation does not constitute a bubble.

The bearish viewpoint was most systematically outlined by Goldman Sachs semiconductor analyst Jim Covello in a report in April.

He noted that nearly all the value in the AI supply chain is flowing to semiconductor companies—an unprecedented and unsustainable phenomenon in history. Chipmakers should benefit when their customers thrive, but in this cycle, their prosperity comes at the expense of the entire upstream supply chain.

NVIDIA's net profit has grown approximately 20-fold since the launch of ChatGPT; major hyperscale cloud providers have exhausted their operating cash flow and are now turning to debt—data center-related debt issuance is projected to reach about $182 billion in 2025, doubling from 2024.

The MIT Nanda study shows that 95% of companies investing in generative AI have seen zero returns. This disconnect may persist for a while, but it cannot last indefinitely.

The discussion also involved a more complex layer: the financial cycle between hyperscale cloud providers and AI labs.

According to corporate disclosure documents compiled by The Information, OpenAI and Anthropic together 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 services backlog, $280 billion is tied to OpenAI;
• Of Oracle's $553 billion pipeline, 54% (approximately $300 billion) is committed by OpenAI;
• Of Google's $467.6 billion, Anthropic accounts for 43% (approximately $200 billion);
• Amazon's corresponding exposure also reaches 51% of its $464 billion backlog.

This financing structure is inherently cyclical. Microsoft’s $13 billion investment in OpenAI is primarily redeemed in the form of Azure credits, which OpenAI uses to purchase Azure computing power, and Microsoft then counts this as cloud revenue.

The same hyperscale cloud service provider is both an equity investor in the AI lab and a service supplier charging for compute resources.

This structure is also reflected in the profit figures. Alphabet reported a record first-quarter profit of $62.6 billion, approximately $28.7 billion—nearly half of which—came from the unrealized gain on its stake in Anthropic.

Of Amazon's $30.3 billion first-quarter profit, $16.8 billion stemmed from pre-tax unrealized gains from Anthropic, while its free cash flow plummeted 95% to $1.2 billion due to $44.2 billion in data center capital expenditures during the same period.

The sustainability of this system depends on AI labs’ ongoing ability to secure external funding to fulfill their cloud computing commitments, which in turn relies on corporate clients’ continued willingness to pay rising token bills.

Reports indicate that Anthropic currently incurs costs of up to $3 for every $1 in revenue. Once the pace of financing slows, the credibility of cloud revenue projections will decline, putting pressure on the valuation multiples of hyperscale cloud providers.

This chain transmits in both directions and will also break in both directions.

The current situation does not constitute a typical bubble scenario.

In terms of valuation multiples, the Big Seven tech companies currently trade at approximately a 20x forward P/E ratio, significantly below the peaks of 52x during the 2000 tech bubble, 67x during the 1989 Japanese market peak, or 34x during the "Nifty Fifty" era.

AI technology itself is real. For heavy user groups, the data on productivity gains is verifiable. OpenAI has an annual revenue of approximately $20 billion, and Anthropic has approximately $4.3 billion—neither lab will disappear.

Today, token cost (computational overhead) has become a critical factor determining the success of AI, whereas just six months ago, few even discussed this topic.

Back then, everyone only cared about whether the technology worked. Now the answer is clear: in specific use cases and among certain groups, the technology indeed works.

But a new question arises: Can downstream companies pass on the savings from AI in time to outpace the valuation window offered by capital markets to AI labs and cloud giants?

Those bullish on AI believe that as the technology continues to mature, companies' ROI will turn positive within 1 to 1.5 years.

Skeptics believe that more executives will follow McDonald’s example, publicly complaining about the low return on AI investments and beginning to cut budgets.

Both possibilities may be occurring simultaneously, and the outcome remains uncertain. The only certainty is that the old myth—that “rising token consumption means AI transformation has succeeded”—has been debunked.

High token consumption does not equate to commercial value; these two bubbles will eventually burst. The AI bill has come due, but it’s still unknown who will pay it.