# Vertex AI Batch Prediction Job

> Vertex AI batch prediction jobs bill for the machine resources held between job start and completion, then terminate. Nothing keeps billing afterwards. Because Cloud Asset Inventory does not index the type at all, ZopNight lists batch predictions from the live aiplatform API across 32 regions and attributes their compute from billing actuals.

Source: https://zop.dev/integrations/gcp/vertex-batch-prediction-job
Updated: 2026-08-19

---

A batch prediction job runs a model against a large dataset offline, billed for the machine resources consumed during the run. Jobs terminate on completion, so cost risk lies in oversized machine types and repeated failed runs.

## Billing that starts and ends with the run

A batch prediction job holds a fleet of machines (node count times machine type, plus any accelerators) for exactly as long as the job takes. When it completes, billing stops; nothing stands idle afterwards. The whole cost question therefore collapses to two decisions made at submission time: the machine specification, and the dataset the job must chew through.

## Why batch predictions need the live API

Cloud Asset Inventory does not index BatchPredictionJob at all, so an inventory built from CAI alone would simply miss the type. ZopNight lists batch prediction jobs from the live aiplatform REST API across 32 Vertex regions and maps them through the same enrichment path as every other Vertex resource, attributing their compute from billing actuals. There is no stop verb to schedule; the job is its own lifecycle. Visibility and sizing are the levers that remain.

## Oversized machines and repeated failures

Waste concentrates in three places: machine types copied from an online-serving configuration that needs headroom batch work does not; failed jobs resubmitted repeatedly against the full dataset instead of a sample; and recurring batch predictions feeding a downstream consumer that no longer exists. Because each run terminates cleanly, none of this ever appears as idle capacity. It appears only in the run history, which is why the history deserves a regular read.

## Batch prediction runs in the console

Google Cloud console → Vertex AI → Batch predictions lists jobs per region with state, duration, and the model used. Recent history is the audit surface: scan for repeats, failures, and machine types out of proportion to the dataset.