AI Compute Costs Overtake Payroll for Some Companies

Axios reported in April 2026 that IT budgets are getting blown out as some companies spend more on AI than on employee salaries, quoting Nvidia vice president of applied deep learning Bryan Catanzaro: "For my team, the cost of compute is far beyond the costs of the employees." Uber is the most concrete example: The Information first reported that the company's CTO revealed Uber had blown through its entire 2026 AI budget in about four months, after the company had encouraged staff to use AI "as much as possible" and tracked usage on internal leaderboards. Bloomberg and TechCrunch reported that Uber responded by capping spending at $1,500 per employee per month on each agentic coding tool, including Anthropic's Claude Code and Cursor, with an internal dashboard to track usage. Fortune separately reported that some engineering teams have reduced Claude Code licenses and shifted toward GitHub Copilot CLI as token costs and usage patterns evolved. Axios cited a projection that worldwide IT spending will reach $6.31 trillion in 2026, with AI infrastructure a major driver.

The common mechanism across these reports is token-based and consumption pricing on hosted models: sustained high-volume usage, prompt-heavy workflows, and many concurrently running coding agents scale linearly or worse with cost, and that spend does not disappear when headcount is unchanged. For teams running custom inference or training workloads, GPU-hour charges, data transfer, and orchestration overhead compound the same dynamic.

The reported responses cluster around a few concrete levers: usage caps (Uber's per-tool, per-employee dollar limit), tool substitution (moving between Claude Code, Copilot, and other agentic coding tools based on relative token cost), and tighter internal telemetry. ML platform teams should treat token and GPU-hour usage as a first-class recurring line item, with cost-aware routing between cheaper and more capable models and rate-limiting for expensive agentic workflows, rather than treating AI tool access as an unmetered productivity perk.

Track vendor pricing changes and enterprise licensing terms that cap or bundle consumption, further usage-limit policies at large engineering organizations, and whether Uber COO Andrew Macdonald's public skepticism about AI's productivity impact, reported by Fortune, spreads to other large adopters. Also watch earnings commentary from cloud and model providers for signs that enterprise cost-consciousness is affecting consumption growth.

These are reported budget and usage decisions at specific companies (Uber, and unnamed firms in Axios's and Fortune's reporting), not evidence of an industry-wide reversal on AI adoption; companies citing high costs are still using the tools, just with tighter controls. The underlying claim that AI is a net cost loser rather than a net productivity gain remains contested and is not established by these reports alone.