Uber, Nvidia warn AI costs exceed payroll

AI was supposed to be the clean spreadsheet story — fewer people, more software, lower costs. But the weird thing happening in 2026 is that some companies are finding the opposite. The labor line is not the one exploding first. The infrastructure line is. Nvidia and Uber executives have now said out loud what a lot of finance teams were already seeing in private: once AI moves from demo to daily use, compute, orchestration, and tool usage can outrun payroll. ### What actually got said? The sharpest line came from Nvidia’s Bryan Catanzaro, a vice president of applied deep learning, who said the cost of compute for his team is “far beyond” the cost of employees. Around the same time, Uber CTO Praveen Neppalli Naga described a version of the same problem from the buyer side — heavy internal use of coding copilots and AI tooling pushed spending much faster than expected. (The Economic Times; Yahoo Finance; TechSpot.) ### Why is this surprising? Because the sales pitch for enterprise AI has mostly been labor substitution. You automate support, coding, document work, QA, operations — and the software should cost less than the humans whose time it replaces. But that only works if usage stays bounded. In practice, once teams get access, prompts multiply, agents call other agents, context windows get bigger, and every “small” task starts carrying token, model, cloud, and monitoring costs. Basically, AI doesn’t behave like a normal SaaS seat. It behaves more like a meter running in the background. (Uber; Nvidia.) ### Why do costs jump so fast? The catch is that the model bill is only one layer. Companies also pay for GPUs, cloud instances, storage, retrieval systems, guardrails, observability, fine-tuning, and the plumbing that routes work between models and internal systems. Nvidia has been pushing “cost per token” as the metric that matters because sticker price on hardware misses the real operating bill. That tells you where the market is already moving — away from “can we use AI?” and toward “can we afford this workload at scale?” (Nvidia.) ### So is AI not worth it? Not exactly. Nvidia’s own 2026 surveys still show companies increasing AI budgets, especially in finance, because plenty of firms are getting revenue gains, productivity gains, or both. The point is narrower and more uncomfortable: ROI is real for some use cases, but not automatic for all of them. A flashy assistant that saves a few minutes a day can be a terrible business if it burns expensive inference all day long. The easy wins are getting separated from the vanity projects. (Nvidia.) ### Why does this connect to layoffs? Because companies are trying to fund two transitions at once. They still carry normal payroll, and now they also carry a fast-growing AI stack. That creates pressure to cut elsewhere. Cloudflare said on May 7 that it was cutting more than 1,100 roles as it restructures for the future, and tech layoff trackers show 2026 job cuts continuing to pile up across the sector. Not every layoff is “because AI,” but AI spending is clearly becoming one more budget line that executives are defending by squeezing others. (Cloudflare; Layoffs.fyi; TechCrunch.) ### What are buyers doing now? They’re getting stricter. Vendors can’t just promise transformation anymore. They need to show replacement math — hours removed, tickets closed, code shipped faster, fraud caught sooner, churn reduced. If an AI feature does not clearly replace expensive human minutes or create measurable revenue, it starts looking like a luxury. That is a big shift from the 2023 and 2024 mood, when adoption alone counted as progress. (Uber; Nvidia.) ### What’s the bottom line? The AI story is maturing. That means less magic, more accounting. The next winners probably won’t be the companies using the most AI. They’ll be the ones that know exactly which AI tasks are worth the bill.