Nvidia makes the chips inside nearly every major AI system built since 2023. In 2026, one of its vice presidents said something that once would have sounded absurd. For his own team, the cost of running AI outweighed what the company paid the employees who use it.
That line captures a shift across corporate America. Companies spent two years racing to put AI into as much work as possible, on the promise that it would pay for itself. Many never settled what the tools should cost or what they had to deliver. The bills that arrived that year told a more complicated story.
In late 2025, Uber gave thousands of engineers broad access to an AI coding assistant, and use spread fast. By spring, its finance team found the whole annual AI budget spent in about four months. Its chief operating officer told analysts the spending was getting harder to justify. Microsoft, one of the AI industry’s biggest backers, canceled most of its internal licenses for Anthropic’s coding tool, Claude Code, partly over cost.
Some of the pressure is built into the pricing. For years, coding assistants such as GitHub Copilot charged a flat monthly fee, however much a developer used them. In April, GitHub announced it would switch to billing based on use, measured by how much text the AI reads and writes. GitHub said the tool had come to do far heavier work than it was first priced for. Once the new billing took effect in June, some developers said they used up a month’s allowance within hours. A $10 or $39 monthly plan could become a bill in the hundreds.
The pattern runs wider than single companies. MIT’s NANDA initiative studied business pilots of generative AI, the kind that writes and creates content, and found that 95% had no measurable effect on profit. The researchers blamed how companies rolled the tools out, not the AI itself. The communications firm Sinch surveyed more than 2,500 senior decision-makers. It found that 74% of their companies had already pulled an AI tool that dealt directly with customers. Most often, the tool had leaked customer data or given an answer that damaged trust.
Yet 98% of those same companies, including the ones pulling tools back, said they still planned to raise AI spending in 2026. The cuts look selective, aimed at the tools that failed to earn back their cost.
