The famous Starbucks company withdrew its smart inventory system from more than eleven thousand coffee shops in the United States, basically because it did not know how to distinguish a whole milk from another type.
Uber ended up recently confessing that it spent the annual budget for artificial intelligence in just four months.
And Microsoft just asked its own engineers to stop using Claude Code because the bill was becoming unpayable.
Is the corporate euphoria over AI over?
By: Gabriel E. Levy B.
The story of Starbucks sounds almost like a black comedy.
In September 2025, the specialty coffee chain proudly announced a deal with NomadGo, a Washington startup founded by a former Microsoft executive.
The promise sounded unbeatable. Baristas would point a tablet at shelves and a computer vision system would count all inventory eight times faster than a human, with ninety-nine percent accuracy.
The company deployed it in more than eleven thousand North American stores, although they realized that from the first day something was wrong.
The company’s own promotional video showed the system failing to recognize a bottle of peppermint syrup while counting neighboring bottles. It was there, in plain sight. They didn’t even bother to hide it, they just thought it was a minor bug that the engineers of the developer company could fix soon.
In January 2026, Reuters published a devastating investigation.
Interviews with baristas, videos of mistakes, leaked photographs.
An employee with thirteen years in the chain, Jake Domey, said that in his store they threw away three entire garbage bags full of food due to overstocking, that is, because the Artificial Intelligence asked for supplies that the stores did not really need.
The most waste Domey saw in his entire career, not even a rookie with days of training would make a mistake of this level.
The reason? Torani syrup bottles are identical with different colored labels.
Whole, skimmed and oat milk come in almost the same white containers.
What in a demonstration with perfect light seemed like science fiction, in the darkness of the cafeteria at seven in the morning, with a barista in a hurry, systematically confused the available products, creating shortages in some products and in other inventory saturation.
On May 18, 2026, an internal bulletin reported that they had made the decision to definitively withdraw the artificial intelligence system that they incorporated to control inventories and in less than three days, the news spread like wildfire in the international press.
The official explanation spoke of standardization and consistency.
The reality, according to the employees themselves, was simpler: the tool never worked.
The other Achilles’ heel is tokens
Before knowing why the bill got out of hand for Uber, Microsoft and the rest, you have to understand a word that almost no one outside the technological world knows: token.
A token is the minimum unit with which artificial intelligence models count the work they do. It’s not exactly a word or a letter. It’s something in between, a little piece of text.
The word hello can be a single token. The word hippopotamus can be split into two or three. A complete sentence, in fifteen or twenty. Every time an engineer asks Claude, ChatGPT, or any other model for something, the tokens that come in (the question) and the tokens that go out (the answer) are counted. And for each one you pay.
The problem is that modern tools, those they call agents, do not respond with a single sentence and are silent. They work alone for hours. They read entire files, try options, correct themselves, start over.
A single query can consume hundreds of thousands of tokens.
Therein lies the trap that no one saw coming. The more useful the tool becomes, the more engineers use it. The more they use it, the more tokens they consume. The more tokens they consume, the higher the bill. The original promise was exactly the opposite. More AI, less costs. But the token math decided to rebel against the Script.
Uber burned a year’s budget in four months
In mid-April, The Information published an interview with Praveen Neppalli Naga, the company’s chief technology officer. What he said left more than one CFO in Silicon Valley cold.
Uber implemented Claude Code, Anthropic’s AI-powered scheduling tool, in December 2025.
By March, eighty-four percent of its 5,000 engineers were using it. Seventy percent of all the code that was uploaded to the system came from machines, not humans. It sounds like a dream of productivity fulfilled.
But then the bills arrived. The average engineer spent between one hundred and fifty to two hundred and fifty dollars a month. The most intensive users consumed between five hundred and two thousand dollars each.
Naga himself spent $1,200 on a single two-hour demo session. Yes, twelve hundred dollars. Two hours.
The phrase he repeated to journalist Laura Bratton sums up the whole thing. He went back to square one because the budget he thought he needed for the whole year was already blown away.
The most curious thing about the case is that Uber had gamified consumption.
There were internal leaderboards that ranked engineers by how much artificial intelligence they used. They literally turned uncontrolled spending into a game with a podium. And then they were shocked to run out of money.
They are not alone.
Meta built an internal panel called Claudeonomics where more than eighty-five thousand employees competed to be the king of token consumption, with nicknames such as Token Legend or Cache Wizard.
The top user burned two hundred and eighty-one billion tokens in thirty days. Neither Mark Zuckerberg nor the CTO made it into the top 250.
Meta closed the panel within two days of being leaked.
Amazon, for its part, pushes its developers to have more than eighty percent use AI tools every week.
Employees called the phenomenon tokenmaxxing and began artificially inflating their consumption just to meet metrics. A study by the firm Jellyfish revealed something devastating. The most intensive users consume ten times more tokens than the average, but they only achieve twice the productivity.
The final blow came from Microsoft
And then, on May 14, came the bombshell that really rocked Silicon Valley. The Verge’s Tom Warren revealed in his newsletter that Microsoft was canceling almost all of Claude Code’s internal licenses for engineers in its Experiences and Devices division. That division is what Windows, Office, Outlook, Teams and Surface do. Thousands of people. The productive heart of the company.
The deadline for migration to Microsoft’s internal tool, GitHub Copilot CLI, is June 30. Just as the fiscal year closes. Pure coincidence.
The juiciest detail of the matter is the corporate paradox. In November 2025, Microsoft had announced an investment of up to five billion dollars in Anthropic, the creator of Claude.
The same Microsoft that put five billion into the company, six months later tells its own employees to stop using its product because it costs too much.
The internal memo signed by Executive Vice President Rajesh Jha sold the decision on arguments of technological unification. Anonymous sources who spoke to The Verge said otherwise. There is a financial component that weighs heavily, and a lot.
What no one wants to acknowledge out loud
Bryan Catanzaro, vice president of applied deep learning at Nvidia, the company that sells the chips that drive all this madness, told Axios on April 26 a phrase that went around the world: For his team, the cost of computing is far above the cost of employees.
Read it again: He’s a senior executive at Nvidia, the company that benefits the most from the artificial intelligence boom, admitting that right now it’s more expensive to use AI than to pay people.
Goldman Sachs, in its May 8 report, projected that global token consumption will increase twenty-fourfold by 2030.
Gartner warned that although the unit cost per token will fall by ninety percent in that same period, the total costs of companies will rise because AI agents consume between five and thirty times more tokens than a normal chatbot.
The big four, Amazon, Microsoft, Alphabet, and Meta, are going to spend seven hundred and twenty-five billion dollars on AI infrastructure in 2026 alone. Seventy-seven percent more than last year.
And meanwhile, at a Connecticut coffee shop, a barista counts syrups with a pencil and notebook.
At Uber’s offices, a chief technology officer returns to the dashboard to recalculate.
At Microsoft, thousands of engineers went back to programming manually because it’s impossible to continue paying massively for AI.
Artificial intelligence is not sinking. Anthropic has a turnover of thirty billion dollars a year, its growth is unstoppable. But the story we were told, that AI was going to cut costs and replace workers without breaking a sweat, is going through its first major reality test. And for the moment, the numbers are not coming out as promised.
In short, Three Hits in a Row Deflate the Promise of Artificial Intelligence. Starbucks retired its smart inventory because it confused milks and syrups. Uber spent the annual AI budget in four months. Microsoft canceled Claude Code licenses to its own engineers because the bill was unpayable. Even an Nvidia VP admitted that using AI costs more than paying humans. The bubble shows cracks.
