AI factories are often cast as inflexible loads that strain the grid. This paper shows the opposite: a properly orchestrated AI factory can sustain computational throughput while simultaneously operating as a dispatchable grid asset. We document a joint proof-of-concept between Uplight and Hammerhead AI demonstrating “Grid-to-GPU” orchestration, the full signal path from utility constraint to GPU-level response, executed on production-grade platforms with real-time telemetry and verified POI compliance. Three co-located assets (a 12 MW AI factory, a solar-plus-storage system, and an industrial park) were operated under a shared 21 MW interconnection limit across three scenarios of escalating grid stress. In two of three, the AI factory absorbed zero load reduction; portfolio-level coordination was sufficient. In the third, Uplight’s platform dispatched a 1.5 MW load reduction that Hammerhead’s ORCA delivered through multi-layer orchestration across power, cooling, and compute rather than a hard shutdown, preserving job continuity while meeting the POI constraint. The implication is concrete: AI factories can meet flexible interconnection requirements today while protecting compute throughput across the majority of operating conditions, provided the orchestration layer reaches from grid signal to GPU scheduler.