Compute Cost-and-Constraint Watch
The price of what the AI buildout buys, and the physical brake on it — the Creativity Watch's third leg (spend = the CapEx Medusa; price and brake = here). Monthly composite of constraint legs: electric-power PPI (the regulated T&D 'who pays' leg), copper (transformer/equipment cost), warehouse-construction PPI (the genre data centers live in), and public frontier-model API price lists (inference cost deflation, via the litellm community price file's git history — timestamped by commit). GPU spot (vast.ai H100-class median $/hr + p90/p10 spread) rides as metadata, accumulating from 2026-10-03. LBNL interconnection queue: DECLARED ABSENT via automated path (403 to bots; parked manual-curate ask). States: CHEAPENING (deflation dominates) / STABLE / CONSTRAINED (the brake is binding). FALSIFIER RESULT, printed full-size: the frozen lead-lag bar vs the estate's own Medusa (BEST_CAPEX revisions) FAILED — rho_lead1 -0.43 vs rho_0 -0.48: the composite is COINCIDENT with the Medusa, not leading. It prints as a DESCRIPTIVE constraint gauge (DeVries negative control 0/18 false fires; the 2026 power-crunch window reproduced); the leading-instrument claim is null.
State History
State Definitions
CHEAPENING — deflation is outrunning the brake (z <= -1.0): compute gets cheaper faster than constraints bind.
CONSTRAINED with composite z >= +2.0.
inference deflation and the physical brake are in balance (|z| < 1.0).
CONSTRAINED — the physical brake is binding (z >= +1.0): power, equipment and construction costs outrun inference deflation.
Sources
Fires Together With
Sensors that were historically elevated on the same days. Higher overlap suggests correlated stress signals.
Read-only intelligence. Sensor states are diagnostic observations, not predictions.