Axios: AI costs exceed salaries
- Uber’s AI spending squeeze became a boardroom issue after reports said some companies now pay more for models, tokens and cloud compute than for staff doing the work. - Uber CTO Praveen Neppalli Naga said the company had already burned through its full 2026 AI budget by April as Claude Code adoption surged across engineering. - Investors are pressing for proof that higher token and infrastructure bills produce revenue or labor savings, not just usage growth. (marketplace.org)
Some companies are now spending more on artificial intelligence tools and computing power than on the employees those tools are meant to assist. (marketplace.org) The pressure point is not just software licenses. It is token charges, cloud bills and infrastructure needed to run systems that answer prompts, write code and complete multi-step tasks. (cnbc.com) (platform.claude.com) Uber has become the clearest recent example. The Information reported last week that Chief Technology Officer Praveen Neppalli Naga said the company had already maxed out its full-year 2026 AI budget just months into the year. (theinformation.com) The tool driving that spike was Anthropic’s Claude Code, according to the report. Secondary coverage said Uber rolled it out in December 2025 and then saw usage jump fast enough to force a budget rethink by April 2026. (theinformation.com) (aimagazine.com) A token is the unit these systems bill against, roughly the chunks of text going in and out of a model. A simple chat uses hundreds of tokens, while an autonomous coding session can consume thousands more in one run. (cnbc.com) Anthropic’s current posted API rates show how that adds up at scale. Claude Opus 4.7 is listed at $5 per million input tokens and $25 per million output tokens, while Sonnet 4.6 is $3 and $15. (platform.claude.com) The spending problem is getting harder to see because many companies still measure adoption by usage. CNBC reported that Meta and Shopify have built internal leaderboards tracking how many tokens employees use. (cnbc.com) That can reward volume instead of results. Databricks Chief Executive Ali Ghodsi told CNBC that if a company wants to burn money, it can simply resubmit the same query repeatedly without producing anything useful. (cnbc.com) Executives are also struggling to prove return on investment. Jen Stave of the Harvard Business School AI Institute told CNBC that chief technology and information officers are having a hard time finding a workable ROI framework. (cnbc.com) That is colliding with a broader financing problem. Data centers take one to two years to build, Anthropic Chief Executive Dario Amodei said, so companies are committing billions before they know whether demand and revenue will match the forecasts. (cnbc.com) The result is a shift from AI as a strategic promise to AI as a line item. Companies can still justify the spend if coding tools cut labor hours or raise output, but boards and investors are asking for that math now, not later. (marketplace.org) (cnbc.com)
