A model that watches a video, reads its captions, and hears its narration must agree with itself about what happened, and that agreement comes from labels and the people who decide what each frame, sentence, and sound clip means. Multimodal foundation models raise the stakes because a single training example carries several modalities that must point to the same interpretation, moving data labeling services from a back-office chore to a determinant of model quality. Stanford's AI Index reports that training datasets double roughly every eight months while compute doubles every five, with image and video work at the center of that growth. Architecture and compute set potential, but annotation sets the ceiling; when labels disagree across modalities, no amount of parameters rescues the result. Single-modality annotation has a forgiving failure mode, but multimodal annotation removes that cushion, as conflicting supervision signals for one moment, such as a caption calling a gesture "friendly" while audio tags it "hostile," teach the model a contradiction that becomes a systematic wobble. Each modality has its own conventions—text annotators think in spans and entities, image annotators in bounding boxes and segmentation masks, audio annotators in timestamps and speaker turns—so a person labeling a cooking video must reconcile the spoken instruction, on-screen action, and caption. Dataset scale magnifies every inconsistency, as a small ambiguity replicated across millions of aligned pairs becomes a stubborn bias. Three workloads dominate demand: pretraining and fine-tuning corpora need vast volumes of aligned pairs, including images with