Companies budgeted for chatbots and deployed agents. Those are different things with different cost profiles, and the bills are arriving now. Token costs for agentic workloads scale with usage, not with headcount. An agent that searches, reasons, writes code, tests it, and loops until a task is done can consume thousands of tokens where a chatbot exchange takes a handful.
Some organizations are discovering this gap at the quarterly earnings level. The conversation has moved from asking what AI can do to what it's actually costing (and whether it's worht it).
AI Compute Costs Are Outpacing Salaries at Some Companies
What Happened
Uber's CTO burned through his full 2026 AI budget on token costs alone before Q2. Bryan Catanzaro, VP of applied deep learning at Nvidia, told Axios that compute already costs more than his team's salaries. A Goldman Sachs survey found large companies overrunning their AI budgets by orders of magnitude. Worldwide IT spending is projected to reach $6.31 trillion in 2026, up 13.5% from 2025, driven by AI infrastructure and cloud services.
Why It Matters
Developers building AI-integrated systems are increasingly accountable for token efficiency. Knowing when a frontier model is worth the cost versus a cheaper alternative, and when an agent loop makes sense versus a simpler pipeline, is becoming a baseline expectation rather than an optimization exercise.
Source: Axios
Tags: AI Tooling, Career, Industry
The Expert Takeaway
So does AI cost more than humans? That's not exactly clear yet. And phrasing the idea that way mostly obscures a more specific thing. Agentic AI workloads consume tokens at a fundamentally different rate than conversational ones.
Companies that modeled their AI spend on a subscription seat price are finding the actual usage pattern is something else. So accountability is popping up now. What are your AI agents producing? Is it worth the line item? I'm starting to hear this on quarterly earnings calls, which is a different conversation than most teams were having twelve months ago.
The engineers who understand how to evaluate which model fits which workload, and how to architect systems that do not burn tokens on tasks that do not require it, are going to be more valuable than the ones who only know how to call an API. For a current view of the tools in play, the best AI tools ranked by real-world usefulness covers the landscape, and AI coding assistants is where the cost-versus-value tradeoff is most visible in practice.
