364 Slides a Shift, Not 222, on the GPU You Already Own: A Deployment Card for Cancer Pathology Foundation Models at Hospital Scale

Cancer pathology foundation models (FMs) are compared almost entirely on area under the curve on retrospective cohorts, where competitive backbones now differ in the third decimal place. That is not the number a pathology department decides on. A department decides on how many slides a card clears in a shift, how many cards it must therefore buy, what the recurring energy bill is, whether the capability survives whichever accelerator procurement supplied, and above all whether protected health information has to leave the building. We argue that the reporting gap, not the accuracy gap, is what currently separates cancer pathology FMs from clinical impact, and we propose a deployment card: five axes, reported jointly on one declared clinical workload, in units a department already plans in. We fill the card in end to end on a real ~12,400-tile clinical slide, across three ViT backbones spanning an order of magnitude in size (DINO ViT-S, UNI, Virchow), on both an NVIDIA H100 and an AMD MI300X, using fused inference operators written once and run unmodified on either vendor. For UNI, one card goes from 222 to 364 slides per 8-hour shift on NVIDIA and 224 to 348 on AMD, energy per slide halves, cost per slide falls by 1.64-2.02x on NVIDIA and 1.55-1.92x on AMD under any electricity rate or amortization schedule, and across four cancer cohorts every |dAUC| is at most 0.001 with mean embedding cosine 0.9998. Read in departmental units the result stops being a benchmark and becomes a procurement fact: a 2,000-slide day that needed ten on-premises cards needs six. It is also a research budget. Continual adaptation after a backbone refresh, and few-shot work on rare cancers and emerging biomarkers, both begin by re-encoding a retrospective archive; for 100,000 slides that falls from 150 GPU-days to 92, and from 2,020 to 1,000 kWh. Weights, numeric precision and diagnostic heads are unchanged by design, so a site can assess the change under its own software change-control and validation procedures rather than as a change to the validated model.