Not long ago, using as much AI as possible was treated as a sign of a good employee. Nvidia's chief executive said that if your highly paid engineer was not burning through a large chunk of that salary in AI usage each year, something was wrong. Meta reportedly ran an internal contest rewarding staff for heavy AI usage. OpenAI's chief executive said in the spring he was excited to see how far this could go.
That mood has broken. Businesses that leaned into heavy AI usage are now looking at bills that grew much faster than the value they got back, and they are pulling back hard.
The clearest number comes from Bain and Company, which advises many large corporations on this exact question. One of its consultants said AI usage costs at some of the firm's enterprise clients have been doubling every couple of months. Take two hundred dollars per developer per month and multiply that across twenty thousand developers, which is common at the large companies Bain works with, and you land on a monthly bill no finance department planned for.
This is not an isolated complaint. A study out of MIT looked at three hundred AI projects across major companies and found that the vast majority delivered no measurable financial return at all. The core problem was not that the AI did not work. It was that companies kept reaching for the most powerful, most expensive AI model for jobs a much cheaper one could have handled just as well, like using a professional camera crew to take a passport photo.
Two of the sharpest public complaints came from executives who are themselves selling AI-adjacent products. Palantir's chief executive said something has gone completely wrong with the token business model, describing American businesses as privately furious about paying heavily for AI systems that create no value while absorbing their company data. Microsoft's chief executive raised a related point: businesses using outside AI tools are paying twice, once in the subscription bill and again by handing over their internal data and work patterns to whichever company built the tool.
The practical fix spreading through large companies now has a name: model routing. Instead of sending every task, from a quick email to a serious research problem, to the most capable and priciest AI model, businesses are building simple rules that send easy work to cheap, fast models and only escalate genuinely hard problems to the expensive ones. Done properly, this alone can cut AI bills by half or more without hurting the quality of the work.
A second force is pushing in the same direction. Chinese AI labs like Moonshot and Zhipu have released open models that perform close to the best American systems at a small fraction of the price, and price-sensitive companies are already switching some of their routine work to them.
For most businesses, the lesson is not that AI stopped being useful. It is that AI spending now needs the same discipline as any other line item: measure what a tool actually returns, match the tool to the size of the job, and stop rewarding usage for its own sake. The companies that figure this out first will spend less and get more, while the ones still chasing token volume as a badge of honor will keep paying for a habit that was never actually productivity.