Vertex AI Workbench Notebook
Does ZopNight manage Vertex AI Workbench Notebook?
Vertex AI Workbench notebooks are managed JupyterLab VMs, often GPU-equipped, that bill continuously while running. ZopNight discovers them via Cloud Asset Inventory, tracks usage over a 42-day lookback, and stops instances on schedule outside working hours, which halts VM and GPU billing while attached disks continue to bill.
Rules that fire on Vertex AI Workbench Notebook
At a glance
| Field | Value |
|---|---|
| Scheduling notes | stops the notebook instance, halting VM and GPU billing; restarts it on schedule |
A Workbench notebook instance is a managed JupyterLab VM, often GPU-equipped, that bills continuously while running. Data scientists leaving notebooks on overnight is one of the most universal ML cost leaks.
A JupyterLab VM that bills like any VM
Under the managed surface, a Workbench instance is a Compute Engine VM: it bills for its machine type and any attached GPU the entire time it runs, regardless of whether a kernel is executing anything. Attached persistent disks bill by provisioned size around the clock (including while the instance is stopped), so a stopped notebook is cheap but not free. GPU-equipped instances concentrate the risk, since the accelerator often costs several times the VM underneath it and idles just as silently.
Scheduled stops for Workbench instances
Workbench notebooks are one of the few Vertex AI types with a real stop verb. ZopNight discovers notebook instances via Cloud Asset Inventory, tracks usage over the 42-day lookback, and stops notebooks on schedule outside working hours, halting VM and GPU billing while preserving the disk and its contents. On the next scheduled start, the instance comes back where it left off. A notebook used during business hours needs the machine for roughly 45 of the week’s 168 hours; the schedule recovers the rest.
The overnight-notebook leak
The classic patterns: a GPU notebook left running over the weekend after a Friday experiment; instances provisioned per person and forgotten when that person changes projects; and machine types chosen for an occasional heavy job, billing at that size through weeks of light editing. Because each instance belongs to one individual, nobody reviews the fleet. That is precisely the gap scheduled stops close.
Workbench instances in the console
Google Cloud console → Vertex AI → Workbench → Instances lists each instance with machine type, GPU, and state. The status column tells you which ones are billing right now.