Vertex AI Data Labeling Job

A data labeling job coordinates human annotation of training datasets, billed per labeled unit. Spend tracks dataset size and task complexity rather than compute.

Per-label pricing, not compute pricing

A labeling job’s meter is the number of units annotated (images, video frames, text segments), scaled by task complexity, with multi-annotator consensus multiplying the count again. No machines are held and nothing bills after the job completes, which makes this the one Vertex job type whose cost conversation is about people rather than hardware. The corollary is predictability: spend scales linearly with the dataset submitted, so the estimate is knowable before launch in a way GPU training rarely is.

Labeling jobs inside ZopNight’s ML view

ZopNight discovers labeling jobs via Cloud Asset Inventory for complete ML pipeline visibility. Annotation is a genuine input cost of a model, and leaving it out of attribution understates what a trained model actually cost to produce. Labeling jobs are not schedulable, and there is nothing to stop; the platform’s contribution is keeping human-labeling spend in the same ledger as the compute it enables.

Controlling annotation spend

Waste wears a different shape here: datasets submitted wholesale when active-learning selection would have labeled a fraction; consensus review applied uniformly, including to easy examples a single annotator handles fine; and re-labeling passes triggered by taxonomy changes that better upfront label design would have avoided. Since each unit costs the same whether or not it improves the model, the sharpest lever is choosing what to label at all.

Labeling tasks in the console

Google Cloud console → Vertex AI → Labeling tasks shows each job with progress and unit counts. Compare labeled volume against what training runs actually consumed. A gap means annotation money outran the models it was meant to feed, and the next labeling budget should shrink to match real consumption.