The story doing the rounds in Silicon Valley
On 28 May 2026, the American outlet Axios (one of the most widely followed business and technology news sites in the United States) published a report that immediately did the rounds across the industry. An AI consultant said that one of his clients had spent half a billion dollars, 500 million, in a single month on Anthropic's Claude assistant. The cause: the company had given its employees unlimited access to the tool, with no spending cap, no quota, no oversight.
The company was not named. But as Tom's Hardware pointed out, the sheer size of the sum narrows the list of suspects down to the very largest global corporations. Few companies in the world could absorb, even by accident, a $500 million bill in 30 days without noticing immediately.
The story seems absurd at first glance. How can you spend that much, by accident, in a month? But when you look more closely at what is happening in large companies in 2026, it is no longer a comic curiosity. It is the symptom of a deeper phenomenon, which even has a name: "tokenmaxxing". And it is now collapsing. This article unpicks what happened, following on from our article on the AI bubble.
How to spend 500 million without meaning to
To understand how this is technically possible, you have to understand how AI is billed to companies. Unlike conventional software (where you pay a fixed subscription per employee per month), generative AI is billed by usage, based on the number of tokens consumed. As a reminder, as we explained in our article on tokens, a token is the basic unit the AI uses to process text (roughly 0.7 of a word in French).
The problem: this billing model is completely unpredictable. An employee asking a simple question consumes a few cents. But an engineer launching "agentic workflows" (automated tasks in which the AI works autonomously over long periods), very long-context prompts, or parallel coding sessions can generate hundreds or even thousands of dollars in costs per day, on their own.
Multiply that by thousands of enthusiastic employees, with no limit configured, and the bill becomes catastrophic. According to the consultant cited by Axios, enthusiastic adoption turned into an "uncontrolled and devastating burn rate". The heaviest users were the first to send the meters through the roof.
Anthropic does offer control tools for businesses: admin dashboards, per-user limits, compliance tools. But these features have to be enabled manually. In this particular case, they clearly were not. The company handed the keys to its thousands of employees and discovered the bill 30 days later. It is the equivalent of giving an entire company a credit card with no limit and saying "have fun".
"Tokenmaxxing": the trend that has just collapsed
The term "tokenmaxxing" refers to the 2024-2025 tendency to maximise AI usage in companies, as if more usage automatically meant more value. The idea was appealing: if AI makes employees more productive, then they should be pushed to use it as much as possible. Some companies even set up internal leaderboards to encourage adoption.
The problem is that this logic completely ignored cost. And 2026 is the year the bill came due. Several concrete cases have emerged in recent weeks, beyond the famous half-billion.
Microsoft. According to The Verge, the company cancelled the majority of its Claude Code licences (Anthropic's coding tool) for employees in several key divisions, partly because of costs. Engineers are being redirected to internal tools or GitHub Copilot by 30 June. The cost per engineer reached 500 to 2,000 dollars a month.
Uber. The company burned through its entire 2026 AI budget in just four months, as early as April. Chief operating officer Andrew Macdonald publicly acknowledged that the costs were becoming "hard to justify". He even admitted that it was very difficult to trace a link between growing use of Claude Code and real improvements for Uber's customers. One of its executives reported spending 1,200 dollars in a single two-hour demo session.
Amazon. According to the Financial Times, the company shut down its internal AI usage leaderboard after employees began generating pointless tasks just to climb the rankings. Some were launching AI agents to carry out tasks of no value whatsoever, purely to inflate their usage statistics, which were used to assess their performance. The result: rising infrastructure costs with no real productivity gain.
Meta. The company also removed the informal "tokenmaxxing" leaderboard its employees had created.
The figure that changes the perspective
To grasp the scale of the shift, one quote bears repeating, from Bryan Catanzaro, vice-president of applied deep learning at Nvidia (the maker of the chips that run virtually all of the world's AI), in an interview with Axios in April 2026: "For my team, the cost of compute far exceeds the cost of employees."
>>Read that sentence carefully. This is not an AI critic speaking. This is the vice-president of the company that sells the chips powering every AI data centre on the planet. When Nvidia's own executive admits that compute costs more than salaries, the "AI will save you money" narrative has a serious problem.
This is probably the most important reversal of 2026 in the corporate world. For two years, AI was sold as a way to cut costs by automating work. We are now discovering that, in many cases, AI costs more than the humans it was supposed to replace or augment.
| Company | What happened |
|---|---|
| Anonymous client | $500 million in 30 days, with no cap configured |
| Microsoft | Cancels most Claude Code licences ($500 to $2,000/engineer/month) |
| Uber | 2026 AI budget exhausted in 4 months |
| Amazon | Shuts down its internal AI usage leaderboard (pointless tasks) |
| Meta | Removes the "tokenmaxxing" leaderboard for its employees |
Why companies got caught out
The underlying mistake is almost universal. Companies treated AI tools like standard software subscriptions ("SaaS", Software as a Service). In that traditional model, you pay a fixed price per user per month, predictable, whatever the usage. An employee who uses Microsoft Word heavily costs the same as one who barely uses it.
Generative AI completely breaks that model. Usage-based billing means costs vary enormously depending on what employees do. And three factors amplify the explosion:
- Choosing the model. Using a high-end model like Claude Opus (which we unpacked in our article on Opus 4.8) costs up to 30 times more than a lightweight model for simple tasks.
- Context length. The more information you feed the AI (whole documents, long conversations), the more each request costs.
- Autonomous agents. The "AI agents" that work in the background, continuously, without human supervision, burn tokens 24/7. At Uber, 11% of technical updates in production were written by agents with no human in the loop.
The result: spending that looked under control turned into a financial black hole, sometimes without executives realising until the bill arrived.
The real problem: return on investment
Beyond the raw costs, the question now nagging finance departments is that of ROI (Return On Investment: does the money spent really bring in more than it costs?).
The Uber case is emblematic. The adoption of Claude Code was a success from a usage standpoint: 32% of engineers were using it in February, 84% in March. 70% of the code shipped came from these tools. From a technical productivity standpoint, it was a triumph. But from a financial standpoint, it was a disaster. And above all, as the chief operating officer admitted, it was impossible to prove that this translates into better products for customers.
This is exactly the tension we anticipated in our article on the AI bubble. Usage is exploding, costs are exploding, but the link to the creation of real value remains unclear. According to Axios, business leaders "are starting to wonder whether the explosion in AI spending is producing meaningful returns".
A troubling detail flagged by Axios: some companies justify layoffs by citing AI's ability to automate jobs. But according to Anuj Kapur, CEO of CloudBees, these workforce reductions may simply be "the only lever they can pull" to offset their AI bills. In other words: humans would be laid off not because AI replaces them effectively, but to pay for the cost of AI itself. A circle whose economic logic deserves to be questioned.
The great U-turn of 2026: AI governance
Faced with these excesses, companies are making a rapid reversal. The watchword is no longer "use AI as much as possible" but "use AI intelligently". In practice, several approaches are becoming widespread.
Spending caps. For the first time, many companies are setting strict monthly limits per employee and per team. Unlimited access is over.
Role-based access. Rather than giving everyone access, AI is reserved for roles where it delivers real value. A senior developer gets access to the powerful models, not everyone.
Choosing the right model. Companies are learning to use lightweight, cheap models for routine tasks, and to reserve top-tier models for complex work. This is exactly the logic we recommended in our article on the environmental cost of AI: the right tool for the right job.
Real-time monitoring. Usage dashboards, automatic alerts, regular audits of token usage. Finance departments are taking back control.
Reusing outputs. Rather than asking AI to generate the same thing over and over, teams are learning to reuse and share what has already been produced.
What this means going forward
This half-a-billion-dollar story is not just an amusing anecdote. It marks a turning point in the maturity of the enterprise AI market. Three consequences are taking shape.
First consequence: the end of naive euphoria. For two years, AI was adopted with sometimes irrational enthusiasm. 2026 is the year of the wake-up call. Companies are moving from unbridled experimentation to disciplined management. That is probably healthy, even if it slows adoption.
Second consequence: pressure on prices. If companies discover that AI costs too much, they will demand lower prices. This is precisely what is pushing Anthropic, OpenAI and Google to cut their rates and offer economy modes, such as Opus 4.8's "Fast mode" at three times less. Commodification is accelerating, as we analysed in our article on the AI bubble.
Third consequence: the question of AI's business model. If even the largest companies struggle to justify the return on investment of AI, what about Anthropic's $965 billion valuation, or the hundreds of billions invested by the tech giants? As we explored, the economic sustainability of the entire AI ecosystem rests on the assumption that this technology creates more value than it costs. That assumption, for the first time, is starting to be seriously questioned.
The lesson for everyone
Whether you are a business leader, a freelancer, or just a user, this story contains a universal lesson about AI in 2026.
AI is not free, and its cost is unpredictable. Unlike conventional software, its price depends on usage, and that usage can explode without anyone noticing. Whether for a company with thousands of employees or a freelancer using APIs, the principle is the same: you have to monitor, cap, and choose the right tool for each task.
For an entrepreneur or a freelancer, the practical lesson is clear. If you use AI through usage-billed APIs (and not just through a fixed subscription like ChatGPT Plus or Claude Pro), keep a close eye on your consumption. A badly configured workflow, an infinite loop, an agent running in the background that has been forgotten, can turn a bill of a few euros into several hundred. At the scale of a company, that becomes half a billion.
The era of "you just have to keep prompting" is over. As one observer summed it up: "it turns out that simply continuing to prompt was not a viable business model, in the end." 2026 is the year when AI in business moved from enthusiastic adolescence to accounting maturity. Probably for the better.
To go further, you can read our article on the AI bubble, which places this story in the wider economic context, our explanation of tokens to understand how billing works, and our article on the environmental cost of AI, which shows that behind every token consumed there is also an ecological bill. AI is transforming the world, but it has a very real price, both financial and physical. And that price, in 2026, is finally starting to be faced head-on.