The price per "token" has fallen. The bill - has exploded. That's the paradox shaking tech companies today: artificial intelligence is getting cheaper per unit, but so addictive in use that costs have run far above anything anyone budgeted for.

The numbers would be almost comical if they weren't true. Uber spent its entire annual AI-coding budget by April. Microsoft took away its engineers' Claude Code licenses just a few months after handing them out. One unnamed company racked up a 500-million-dollar bill to Claude because it forgot to set a limit. A single engineer - 40,000 dollars in tokens in one month.

Behind it all stands a figure that explains everything: spending per engineer has grown around 18.6 times in nine months. A study of 20,000 programmers showed that output is rising, but so are errors and reworks. Jellyfish's research found that the biggest token spenders are twice as productive - but spend ten times more to achieve it. Does it pay off? Nobody knows, because, as one of the researchers admits, most companies still can't measure how much the delivered code is actually worth.

People from the industry talk in language once reserved for addictions. The CFO of Priceline openly compared the situation to dependency and introduced limits on certain teams. The head of enterprise at OpenAI says the conversations are no longer about what AI can do, but "we're spending this much, what visibility do you have?" The translation is clear: the "run faster" euphoria has turned into "hit the brakes."

That's why a whole industry is springing up around the problem - the Linux Foundation founded a standards body called Tokenomics, companies are emerging that measure and optimize spending, existing platforms are adding cost controls. And a Goldman Sachs forecast says global token usage will multiply 24 times by 2030. In other words: the bill that's shocking companies today is only just beginning.