GPU instance auto-shutdown for AWS and Google Cloud
Last updated: August 31, 2026 · Content reviewed quarterly
Idlefy keeps GPU virtual machines on AWS and Google Cloud stopped by default. An engineer leases the machine for a training run, a fine-tune, or an interactive session — from Slack, Telegram, or the web dashboard — and the VM stops itself when the lease expires. The timer is always set by the person who booked the machine; Idlefy never infers a shutdown time from how busy the machine looks. The stop is the cloud's normal power-off, never a terminate: persistent disks and the checkpoints on them survive, and the same machine starts again on the next lease. On machines that can list for tens of dollars an hour, the difference between “stopped by default” and “someone remembers to stop it” is often the largest line on the bill.
An idle GPU is the most expensive idle hour
A general-purpose dev box left on overnight wastes a few dollars. A GPU box left on after the last training run of the day is a different order of magnitude. AWS and Google Cloud list-price examples — what the machine bills running 24/7, and how much of that falls outside a 10×5 work week:
| Instance | GPUs | On-demand (instance) | Running 24/7 | Nights + weekends share |
|---|---|---|---|---|
| g4dn.xlarge AWS | 1 | $0.526/hr | $384/mo | $270/mo est. |
| g6.xlarge AWS | 1 | $0.805/hr | $588/mo | $413/mo est. |
| g6e.xlarge AWS | 1 | $1.861/hr | $1,359/mo | $954/mo est. |
| p5.48xlarge AWS | 8 | $55.040/hr | $40,179/mo | $28,221/mo est. |
| g2-standard-4 GCP | 1 | $0.707/hr | $516/mo | $362/mo est. |
| a3-highgpu-8g GCP | 8 | $87.832/hr | $64,118/mo | $45,035/mo est. |
On-demand list prices, cheapest listed region. Prices as of 2026-08-31. Estimates only — your regions, usage, and discounts will differ. Idlefy manages every GPU family the clouds offer — P- and G-series on AWS, A2/A3/G2 on Google Cloud. Spot instances are supported too, with the obvious caveat: a spot reclaim follows the cloud's own interruption rules, not the lease timer. Per-type detail lives on the instance idle-cost pages, and a different work cadence prices out in the schedule savings calculator.
Compare GPU prices across clouds — on-demand and spot, per GPU.
Off by default, held by lease
With Idlefy, a GPU machine stays stopped until someone takes a lease — a fixed booking of an hour, a workday, or up to 72 hours on Pro, started from Slack, Telegram, or the web dashboard. The machine starts in about a minute, and when the lease expires it stops itself. Managed notebook platforms ship idle timeouts for their own notebooks; a lease covers any GPU VM — training boxes, self-hosted notebooks, inference sandboxes — with an explicit stop time instead of an inferred one.
For a team sharing one expensive box, the lease doubles as visibility: every lease records who has the machine and until when, so “is anyone using the GPU server?” is a glance at the dashboard instead of a message to the group chat. And because every start and stop is logged, the month-end conversation about the GPU bill has receipts.
Working late? Warnings go out 20 and 5 minutes before expiry, and extending is one tap from the same chat. On the Pro plan, the Safety System adds a backstop: if the machine still looks actively in use at expiry and you haven't responded, it grants a single 30-minute grace period and notifies you again — built for the moments you're away from the keyboard mid-epoch. Activity can postpone the stop by that one window; it can never cancel it.
What forgetting costs without this: the $5,000 weekend — the story of one GPU environment left running from Friday to Monday.
Checkpoints, disks, and the one caveat that matters
A lease expiry performs a stop, never a terminate. Root and attached persistent disks (EBS, GCP persistent disks) survive: your environment, datasets on persistent volumes, and checkpoints written there are all present on the next start. The instance keeps its ID, type, and tags — the same machine comes back, not a new one. Compute billing stops with the machine; disk storage continues at the provider's normal rate.
The caveat: most GPU and ML instance families (the p-, g- and a-series included) commonly carry local instance-store scratch drives (on Google Cloud they are attached rather than built in), and data on scratch does not survive a stop. If your training loop writes checkpoints to local scratch for speed, sync them to a persistent disk or object storage before the lease ends. The 20-minute warning exists for exactly this.
Access is deliberately narrow: no agent to install, and on AWS Idlefy can only start or
stop instances tagged idlefy=enabled — a
boundary written into the IAM policy itself. On Google Cloud the same boundary is
enforced in Idlefy's application layer rather than by IAM, an honest difference the
how-it-works page
documents along with the full lease lifecycle.
Common questions
Do stopped GPU instances still cost money?
A stopped instance stops billing for compute — the GPU-hour charge, which is most of the price, ends with the stop. Attached persistent disks and any reserved static IPs continue to bill at the provider's normal rates while the machine is off.
Can an auto-shutdown kill my training run mid-job?
Not by surprise. The stop time is the lease you set yourself — never a guess made from machine activity — with warnings 20 and 5 minutes ahead and one-tap extension from Slack, Telegram, or web. On Pro, the Safety System can add a single 30-minute grace period if the machine still looks in use and you haven't responded.
How much does an idle GPU instance cost?
At list prices, a single-GPU box like g6e.xlarge or g2-standard-4 runs from several hundred to over a thousand dollars a month 24/7, and a multi-GPU machine like p5.48xlarge or a3-highgpu-8g to five figures — the table above prices the share that falls outside working hours. The free Idle Audit reads 30 days of your actual usage and prices every idle hour in your own account.
More edge cases are covered in the FAQ.
Price your idle GPU hours first
The free Idle Audit reads 30 days of usage across your fleet — GPU machines included — and prices every idle hour before you change anything. Read-only, no plan required.
Run a free idle auditAlso running ordinary dev boxes? See auto-shutdown for dev environments, or compare plans on the pricing page.