AI Reliance: a reproducible measure of how deep, and how reversible, systematic AI dependency is

This working paper measures how deep systematic AI dependency now runs and how reversible it is, using only sources that the dependents and providers publish themselves. It is the demand-side companion to "The AI Capital-Infrastructure Barbell," meeting that supply-side analysis at shared chokepoints (a few clouds, one accelerator vendor, a handful of frontier labs). The approach is structural rather than predictive, and roughly half of the evidence pushes against the alarm case and is reported as found. Four data layers are assembled. Reliability: provider incidents are frequent but short and well managed, yet a single uptime number misses the dominant risks, correlated cross-provider failure and the policy off-switch (the June 2026 model export suspension is the worked example). Downstream attribution: products that publicly blame an upstream model for their own incidents do so for roughly 30 percent of incidents across the transparent panel. Penetration: embedding is deepest where exit is hardest, with US federal AI use cases rising about fivefold in two years (OMB inventory, a floor since the Pentagon and Intelligence Community are exempt from reporting). Reversibility: bimodal, reversible on a three-to-six-month clock in the private sector (convergent industry evidence plus two non-survey anchors), far slower and costlier in government, where the 2026 defense procurement dispute shows that vendor-switching is achievable but painful. A reproduce script (Python standard library only) recomputes the incident metrics from raw status-page JSON, and CSV layers carry the data. Every figure is dated and graded by source tier (primary, corroborating, or direction-only). Conflict of interest: this assessment was produced with the assistance of an AI model made by Anthropic, which is one of the providers measured here. Lines where Anthropic is the subject are flagged in the paper, and third-party data is weighted over self-report wherever both are available. The author is independent and holds no position in the companies named, and set the question, method, and interpretation. This is open-science documentation and macro-structural research only, not investment advice.