# Scheduling Vertex AI Workbench Notebook

> Stopping a Vertex AI notebook instance halts VM and accelerator billing together, while the persistent disk keeps both the work and a small charge. Accelerator billing is the line that matters, since an attached A100 bills roughly 10 times the VM under it. Running kernels die at the cutoff, and GPU capacity is re-requested at start.

Source: https://zop.dev/integrations/gcp/vertex-notebook/schedule
Updated: 2026-08-19

---

## The GPU is the reason this schedule exists

A notebook VM alone costs little; the accelerator attached to it does not. An idle A100
burning through nights and weekends is among the most expensive forms of forgetting in cloud
ML, and the scheduled stop addresses precisely that: VM and GPU charges end together at the
evening cron. The disk-only residual that remains is orders of magnitude smaller.

## Work survives on the disk

The instance's persistent disk carries notebooks, checkouts and downloaded datasets across
the stop, billing its provisioned gigabytes throughout. What does not survive is anything
alive: running kernels, in-memory tensors, an unattended training loop mid-epoch. The
schedule quietly enforces the discipline every ML platform team preaches. Long training
belongs in submitted Vertex training jobs, which run on their own compute and their own
billing, unaffected by the notebook's calendar.

## Accelerator capacity is re-negotiated each morning

Stopping releases the GPU back to the regional pool, and the morning start requests one anew.
For plentiful accelerator types this is a formality; for the scarce ones it occasionally is
not, and a start can fail on capacity. Researchers holding hard-to-get accelerator types
should weigh that risk consciously; for everyone else the overnight release is close to
free.

## Idle shutdown catches some of this, the calendar catches the rest

Vertex notebooks support idle-based auto-shutdown, and it should be on. Its weakness is the
work that defeats idleness detection, such as a cell polling a slow query or a browser tab
auto-executing, which can hold a GPU hostage until Monday. A hard weekday-evening and
weekend stop caps the damage at hours instead of days.

## The honest arithmetic

Count accelerator-hours, not instance-hours. A single scheduled notebook with a premium GPU
dark from 19:00 to 07:00 plus weekends avoids roughly 118 accelerator-hours a week, often
more than the rest of a small team's entire compute bill.