Forgot to set limits: one company burned $500 million on Claude in a single month

Author: Bao Yilong; Source: Wall Street Insights

The corporate AI boom is encountering its first truly meaningful billing crisis.

On May 28, according to Axios citing an AI consultant, a corporate client under its umbrella recently spent $500 million in a single month on Claude, solely because no usage limits were set for employees.

Analysis suggests that many companies, when rapidly deploying AI tools, focus on features and promotion but neglect the establishment of cost control mechanisms.

Tech giants like Microsoft and Amazon are taking action to cut internal AI tools or halt projects tracking AI usage to curb what is called "tokenmaxxing" — excessive consumption of tokens.

An Amazon senior vice president had to warn employees:

Please don't use AI just for the sake of using AI.

The core issue facing the market now is no longer "whether to embrace AI," but "after spending so much money, what have we actually gained?"

Amazon's shutdown, internal "score boosting" triggers real costs

Amazon's case reveals another perspective on the difficulties of corporate AI governance.

According to reports citing two insiders, Amazon's developer platform Kiro once had an internal leaderboard called "Kirorank," which scored employees based on their AI activity levels.

However, this leaderboard unexpectedly prompted employees to perform meaningless tasks via AI agents to boost their rankings, directly leading to increased company computing power consumption.

This week, Amazon senior vice president Dave Treadwell admitted to employees that while the leaderboard's initial intent was good, the end result was employees pushing up costs through "tokenmaxxing."

He explicitly instructed staff not to focus on token consumption but to concentrate on building better products, emphasizing: "Don't use AI just for the sake of using AI."

Amazon later confirmed in a statement that this beta dashboard "was not an official or approved tool and has been taken offline."

Meta experienced a similar situation, with employees also attempting to increase token consumption to improve their internal rankings.

This phenomenon indicates that when companies incorporate AI usage into performance assessments, it can backfire, distorting employee incentives into ineffective consumption of computing resources.

Subsequently, Amazon shifted to using "normalization deployment" metrics instead of token consumption, focusing on whether engineers can continuously generate practically valuable code with AI.

It is noteworthy that Amazon's capital expenditure this year is expected to reach $200 billion, with the vast majority flowing into AI and data center infrastructure.

Four core issues: why AI spending hasn't yielded returns

According to Axios, there are four structural obstacles currently facing enterprise AI adoption.

Misaligned use cases. Sophia Velastegui, CEO of Velastegui Ventures and former Microsoft Chief AI Officer, states that most people tend to use AI to automate tasks they dislike rather than those that are most valuable to the company.

She believes companies should focus AI resources on scenarios that directly drive revenue, rather than deploying blindly.

Lack of cost control. AI queries are not free; enterprise plans charge per token, and even simple daily queries can quickly accumulate significant expenses, yet most business units lack clear awareness of this.

People are the biggest bottleneck. Velastegui characterizes the widespread "scattershot" AI authorization approach as a path that fails to deliver substantial returns.

Companies pile numerous AI tools onto employees but lack effective guidance and focus, resulting in low actual adoption efficiency.

Data openness concerns. Josh Pantony, CEO of Boosted.ai, which focuses on AI tools for the financial industry, points out that when companies hesitate to open internal proprietary data to AI agents due to security concerns, the actual effectiveness of these agents is greatly diminished, making ROI impossible to realize.

Token economics: the new core variable in AI narratives

Behind this debate is a more complex investment logic being reshaped.

Wall Street Insights reports that, according to Rich Privorotsky, head of Goldman Sachs' One-Delta division, the core variable in AI trading has shifted from "whether the technology is feasible" to "whether the costs are affordable."

DeepSeek has reportedly cut token prices by 75%, and Xiaomi's MiMo has reduced prices by nearly 99%. This cost compression could trigger a "price war" similar to subsidy competitions.

He notes that infrastructure bottlenecks will eventually ease, and the market should not pay excessive premiums for "problems that are about to be solved."

Privorotsky further hypothesizes whether cheaper tokens will first replace high-cost inference services. If demand expansion lags, revenue growth for cloud providers, model companies, and AI infrastructure could face temporary pressure.

He believes that rationalizing token expenditure may become a key board-level issue in Q2 and Q3 of this year, as important as the AI growth narrative itself.

According to Bloomberg's Silicon Data LLM Token Expenditure Index, token prices have risen about 65% since late February, and the price of AI software in the U.S. has increased by 20% to 37% over the past year.

This cost trend is prompting companies to reevaluate AI procurement strategies. As "getting 90% of the output at 10% of the cost" becomes increasingly feasible, reliance on high-cost cutting-edge models may decline systematically.

Ali Ansari, CEO of AI model training company Micro1, states that companies are experiencing a "healthy oscillation" from overusing AI to using it rationally. He believes:

The only truly effective area for AI right now is programming.

Bull vs. Bear: same reality, two interpretations

Regarding AI investment returns, the same data points to very different conclusions depending on the analytical framework.

The bullish perspective sees the current chaos as normal growing pains during transformation.

According to Goldman Sachs' Jim Schneider in early May, by 2030, agent-based AI will drive token consumption up 24 times, and large-scale cloud service providers and model vendors will see their gross margins turn positive within the next 3 to 12 months.

JPMorgan's economic research also found that by early 2026, Python package usage on PyPI experienced a leap, a trend not seen when ChatGPT launched in 2022, indicating real productivity gains are occurring.

The bearish view, as outlined by Goldman Sachs semiconductor analyst Jim Covello in a April report, states that nearly all value in the AI supply chain flows to semiconductor companies, which is unprecedented and unsustainable. Chips should benefit when customers benefit, but this cycle's prosperity is at the expense of upstream industry consumption.

Both narratives are unfolding simultaneously, and the outcome remains uncertain. What is clear is that the simple equation "token consumption growth equals AI transformation success" has been broken.

From the extreme case of burning $500 million in a single month to Amazon halting leaderboard score boosting, AI investments are under more rigorous return scrutiny. How much real value the next AI bill can generate will be the true verdict of this high-stakes gamble.