The AI euphoria that swept Silicon Valley in recent years has collided with harsh economic reality: major tech companies are increasingly admitting that spending on tokens – the basic unit of AI usage – has spiraled out of control.

Amazon shut down an internal contest encouraging maximum AI use at work, reminding employees: "Please don't use AI just to use AI." Uber imposed a $1,500 monthly token limit for staff after exhausting its annual AI budget early in the year. Even more telling, model developers themselves now acknowledge the scale of the problem: at a recent event, OpenAI CEO Sam Altman admitted that token consumption has become "a huge problem" for businesses promised productivity gains from AI adoption.

This marks a sharp reversal from the rhetoric of just months ago, when the industry promoted "tokenmaxing": the more an employee uses AI, the more valuable they become to their employer. Yet AI has always been expensive, and costs for training and inferencing new models continue to rise. The situation is compounded by the boom in AI agents – systems capable of autonomous operation: according to an April preprint, they consume 1,000 times more tokens than standard chatbots.

All these costs ultimately fall on companies and end users. Unsurprisingly, GitHub's shift to token-based pricing has drawn criticism. Tech giants urgently need a way to offer AI to the masses without astronomical bills – or users will migrate to free, open-source alternatives.

In response to rising costs, some companies are betting on edge computing: Microsoft and Google have unveiled products where models run locally on devices rather than in the cloud. Of course, powerful models like GPT-5 can't run directly on laptops, but for most everyday tasks, lighter versions suffice – and save money on tokens.

However, investments in edge computing remain dwarfed by spending on data centers: cloud infrastructure still underpins both companies' business models. Yet the very turn toward local processing signals recognition that the price of massive AI models simply doesn't justify the pressure on most users. Ultimately, AI's future hinges on how affordable it becomes for the average person. If the price tag proves too high, human labor may once again prove more cost-effective.