Nvidia exec flags compute costs
- Nvidia vice president Bryan Catanzaro said his team’s compute bill now runs “far beyond” payroll, undercutting the idea that AI is already cheaper than staff. - The remark came in an Axios report on April 26, as companies including Uber said heavy model use had already blown through 2026 AI budgets. - Big Tech still plans nearly $700 billion in 2026 AI capex despite unresolved return-on-investment questions. (cnbc.com)
Nvidia vice president Bryan Catanzaro said his team’s compute costs are now “far beyond” employee costs, a blunt measure of how expensive AI use remains inside the industry. (axios.com) Catanzaro leads applied deep learning at Nvidia, the company whose chips power much of the current AI boom. His comment landed in an Axios report published April 26, 2026, about companies spending more on AI systems than on salaries. (axios.com) “Compute” is the rented or owned processing power needed to train and run models, usually in data centers packed with graphics processing units, or GPUs. Those costs rise with every query, every generated answer, and every software agent left running in the background. (axios.com) (fortune.com) The same Axios report said Uber had already exhausted its planned 2026 AI budget, with token costs from heavy model use driving the overrun. That is a different problem from buying chips once; it is an operating expense that keeps recurring as usage grows. (axios.com) That tension sits beside a separate spending surge in infrastructure. CNBC reported in February that Amazon, Microsoft, Alphabet and Meta were on track for close to $700 billion in combined 2026 capital expenditures, largely tied to artificial intelligence. (cnbc.com) The economics matter because many executives sold AI as a labor-saving tool. If model usage costs more than the wages attached to a task, the case for replacing people gets weaker unless the software also lifts output, quality, or speed enough to justify the bill. (axios.com) That skepticism is not new. A 2024 Massachusetts Institute of Technology study on computer-vision tasks found that a large share of work identified as technically automatable was not yet economical to automate. (news.mit.edu) (ide.mit.edu) Axios also reported that AI labs face their own margin pressure because more customer usage can mean more compute spending. That turns growth into a cost problem as well as a revenue story. (axios.com) Nvidia’s warning does not mean companies are pulling back from AI. It means the argument has shifted from “can we deploy it” to “which jobs and workflows produce enough value to cover the meter.” (axios.com) (cnbc.com)
