# Scheduling Compute Engine Instance

> Stopping a Compute Engine instance releases vCPU and memory billing while persistent disks ride through untouched. Local SSDs must have their data discarded to stop, and sustained-use discounts shrink as running hours drop, so the realized saving lands a few percent under what the raw 168-hour week suggests.

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

---

## Cores released, disks kept

The scheduled stop powers the instance down and ends its vCPU and memory charges; every
persistent disk stays attached, intact and billing at its provisioned size. Start returns the
same instance, same name and same internal IP inside its VPC, with the boot disk exactly as it
was left.

## Local SSDs are the disqualifier

An instance with local SSDs cannot simply stop: the stop requires discarding local SSD
contents, because the hardware is physically bound to the host being vacated. Workloads using
local SSD as scratch space can accept the nightly discard; anything treating it as real
storage makes the instance a poor scheduling candidate until the data moves to persistent
disk.

## Sustained use discounts claw back part of the win

GCP automatically discounts instances the longer they run within a month, up to a meaningful
percentage for full-month usage. Scheduling cuts running hours, which also cuts the discount
tier, so an instance stopped 70% of the time saves somewhat less than 70% of its former
bill. The effect is a few percentage points, not a reversal, but forecasts that ignore it
consistently overpromise.

## Committed use makes the schedule pointless

Instances covered by committed use discounts bill the commitment regardless of state. A
scheduled stop on covered vCPUs saves nothing; the commitment invoices on. Check coverage
before scheduling, and aim schedules at the on-demand remainder of the fleet where every
stopped hour is a real saving.

## Addresses on the restart path

An ephemeral external IP is released at stop and replaced at start, breaking any allowlist or
DNS record that memorized it. Reserving a static address fixes the churn, and a reserved
address on a stopped instance bills a small hourly rate, one more line in the honest
overnight residual alongside the disks.

Attached GPUs stop billing with the instance, which makes accelerator-carrying machines the
highest-yield candidates in most projects. They are also the ones where restart capacity deserves
a moment of thought.