AI spending is becoming harder to explain even as model prices fall, CloudZero founder and CTO Erik Peterson argued in a QCon AI presentation. Provider bills often reduce activity to token totals, leaving engineering leaders unable to connect charges to a feature, customer outcome or useful unit of work.
Peterson criticizes “tokenmaxxing,” the practice of treating higher AI consumption as evidence that employees or teams are adopting the technology successfully. More prompts, generated code and automated tasks can increase activity without proving better products, faster delivery or stronger margins. Cheaper tokens may also encourage enough additional use that total spending still rises.
The accounting problem is broader than input and output tokens. Agent retries, hidden reasoning, model routing, audio or image units and vendor-specific formulas can make two apparently similar requests cost different amounts. Conventional cloud bills are imperfect, but they generally expose resources such as compute and storage more clearly than a single aggregate token number.
The talk presents a vendor executive’s analysis and forecasts, not an audited industry-wide spending study. Its practical recommendation is still testable: attribute AI costs to products and workflows, then compare them with business outcomes. A leaderboard based only on consumption can reward waste and hide whether automation is creating value.