# Stop Counting AI Tokens, Start Counting What Ships Shift focus from token consumption to the efficiency of delivered software features. - **Source:** [MIT Technology Review](https://www.technologyreview.com) - **Published:** 2026-11-02 - **Topic:** [Software & Engineering](https://brifing.si/c/engineering) - **Original article:** - **On Brifing:** https://brifing.si/s/0532cc81-9d1e-4247-a728-4f023f9df557 ## The Token Consumption Trap AI coding agents perform iterative tasks like planning, generating, and testing code, which consumes 10 to 50 times more tokens than simple chat prompts. This high usage creates significant financial pressure, as seen when companies like Uber reportedly exhausted their 2026 AI budgets by April due to expanded Claude Code usage. ## Rationing Versus Value Creation Organizations often respond to rising costs by implementing usage caps, seat limits, or strict approval workflows. However, these blanket restrictions can inadvertently stifle the very workflows that deliver the most value. Measuring cost per token only tracks consumption, failing to distinguish between wasteful runaway loops and the actual shipping of software. ## Redefining Success Metrics A superior metric is the cost per merged pull request, calculated by dividing total AI-assisted development costs by the number of changes that pass review. By tracking rework and defects alongside this figure, engineering leaders can ensure that lower costs do not mask declining code quality or increased technical debt. ## Optimizing Inference Workflows Not every coding task requires a frontier model. Organizations should reserve premium models for complex reasoning and novel planning, while routing routine tasks like code generation and refactoring to open-weight models. A gateway can manage this routing based on task complexity, data sensitivity, and latency to optimize overall inference spending. ## Economics of Owned Capacity Rented inference models increase costs with every call, whereas owned capacity allows costs to become more efficient as usage grows. While ownership requires upfront investment in power, maintenance, and operations, it provides control over proprietary code. This approach keeps sensitive repositories secure while enabling larger-scale AI adoption across the enterprise. ## Hybrid Infrastructure Results HPE utilizes a hybrid AI platform to route routine workloads from over 150 internal applications to on-premises open-weight models. This strategy reduced token costs by 30 to 50 percent. By combining owned capacity for high-volume tasks with frontier APIs for demanding work, organizations maintain performance while significantly lowering their total expenditure. ## Identifying Power User Impact Usage data reveals that a small minority of power users often drives the majority of consumption. In some cases, the heaviest users consume 450 times more than the lightest. If these power users become half the workforce, monthly spending can rise 4.4 times, highlighting the necessity of tracking cost per active developer. ## Building Sustainable AI Adoption The goal is to build an operating model where increased AI usage improves development economics rather than undermining them. Leaders should baseline costs before changing infrastructure and test full development cycles on owned capacity. Success is measured by the ability to ship more software efficiently without rationing access to essential tools. ## Sources covering this story - https://www.technologyreview.com/2026/11/02/1147702/stop-counting-tokens-start-counting-what-ships/ — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._